Can AI Solve the Diversity Problem in the Tech Industry? Mitigating

Can AI Solve the Diversity Problem in the

Tech Industry? Mitigating Noise and Bias in

Employment Decision-Making

Kimberly A. Houser*

22 STAN. TECH. L. REV. 290 (2019)

ABSTRACT

After the first diversity report was issued in 2014 revealing the dearth

of women in the tech industry, companies rushed to hire consultants to provide

unconscious bias training to their employees. Unfortunately, recent diversity

reports show no significant improvement, and, in fact, women lost ground

during some of the years. According to a Human Capital Institute survey,

nearly 80% of leaders were still using gut feeling and personal opinion to

make decisions that affected talent-management practices. By incorporating

AI into employment decisions, we can mitigate unconscious bias and

variability (noise) in human decision-making. While some scholars have

warned that using artificial intelligence (AI) in decision-making creates

discriminatory results, they downplay the reason for such occurrences—

humans. The main concerns noted relate to the risk of reproducing bias in an

algorithmic outcome (“garbage in, garbage out”) and the inability to detect

bias due to the lack of understanding of the reason for the algorithmic

outcome (“black box” problem). In this paper, I argue that responsible AI will

abate the problems caused by unconscious biases and noise in human decision-

making, and in doing so increase the hiring, promotion, and retention of

women in the tech industry. The new solutions to the garbage in, garbage out

and black box concerns will be explored. The question is not whether AI should

* Kimberly A. Houser is an assistant professor at Oklahoma State University. The

author would like to thank the participants at the 2018 Law and Ethics of Big Data

Colloquium in Wellesley, Massachusetts, sponsored by Babson College, Virginia Tech,

Center for Business Intelligence and Analytics, Pamplin College of Business, and Indiana

University, Department of Legal Studies, for their helpful remarks. The author would also

like to additionally thank Angie Raymond, Ramesh Sharda, Griffin Pivateau, and Laurie

Swanson Oberhelman for their insightful comments and to Haley Amster, Justin Bryant,

Abigail Pace, Katherine Worden, Collin Hong, and Caroline Lebel for their thoughtful and

thorough editing, proof-reading, and cite-checking.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 291

be incorporated into decisions impacting employment, but rather why in 2019

are we still relying on faulty human decision-making.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 292

TABLE OF CONTENTS

I. INTRODUCTION ........................................................................................................... 292

II. WOMEN IN THE TECH INDUSTRY .............................................................................. 296

A. Issues Faced by Women ............................................................................................. 297

B. Reasons Attributed to the Lack of Women in the Tech Industry.......... 301

III. GENDER DISCRIMINATION LAW ................................................................................ 305

A. Unconscious Bias in Case Law ................................................................................ 307

B. Unconscious Bias in Cases That Were Settled ................................................ 308

IV. THE BUSINESS CASE FOR WOMEN IN TECH ............................................................. 310

A. Financial Benefit ........................................................................................................... 311

B. Increased Numbers of Women in Tech .............................................................. 312

C. Benefits to Women in Leadership ......................................................................... 313

D. The Need to Fill Tech Jobs in 2020 ....................................................................... 314

V. CURRENT DIVERSITY AND INCLUSION METHODS DO NOT WORK ........................ 315

A. Training Does Not Work ........................................................................................... 315

B. Mentoring Programs Do Not Work ..................................................................... 317

VI. UNCONSCIOUS BIAS AND NOISE ................................................................................ 318

A. Unconscious Bias ........................................................................................................... 319

B. Noise..................................................................................................................................... 323

VII. USING AI TO REDUCE BIAS/NOISE IN HUMAN DECISION-MAKING ................ 324

A. Tackling Unconscious Bias ....................................................................................... 324

B. Reducing/Eliminating Noise ................................................................................... 330

VIII. USING AI TO REDUCE ALGORITHMIC BIAS ......................................................... 332

A. “Garbage In, Garbage Out”....................................................................................... 333

B. “Black Box” ....................................................................................................................... 340

IX. LEGAL CONCERNS IN INCORPORATING AI INTO YOUR D&I PLAN ........................ 345

X. CONCLUSION................................................................................................................ 351

I.INTRODUCTION

Although 1.4 million computer science jobs in the United States will be

available by 2020, only 29% of those positions are expected to be filled, and

less than 3% of those jobs will be filled by women.

The New Yorker has

reported that Silicon Valley loses more than $16 billion annually from the

turnover of half of the women who enter the tech field.

This mass exodus

. Swati Mylavarapu, The Lack of Women in Tech Is More Than a Pipeline Problem,

TECHCRUNCH (May 10, 2016), perma.cc/E2KY-PHW2.

. Alex Hickey, Systemic Gender Discrimination Costing Tech Billions, CIODIVE

(Dec. 7, 2017), perma.cc/9BXS-YMEQ; see also Jennifer L. Glass et al., What’s So Special

about STEM? A Comparison of Women's Retention in STEM and Professional Occupations,

92 SOC. FORCES 723 (2013) (demonstrating that women in STEM are significantly more

likely to leave their field than women in other professions).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 293

signals a significant problem in the industry and represents a substantial

obstacle to the U.S. tech industry remaining at the forefront of the world

economy. While tech companies in recent years have spoken about reducing

the gender gap,

little progress has been made.

Traditional methods of hiring include on-campus interviews, online job

postings and referrals.

Applicants who come from referrals are considered

to be better risks.

This type of preference can lead to the exclusion of

qualified candidates and reinforces the homogenization of an organization.

Research has shown that unconscious biases are rife in the tech industry

and one of the main factors negatively impacting women in this field.

According to a Human Capital Institute survey, nearly 80% of leaders were

still using gut feeling and personal opinions to make employment

decisions.

Not only are human decision-makers unaware of their biases,

they are also unaware of the inconsistency of their decisions (known as

noise). As Nobel Prize winner Daniel Kahneman points out, human decision-

making is fraught with bias and unjustifiable variability.

These types of

unconscious biases are linked to discriminatory behavior.

. Thomas Ricker, How Do Tech’s Biggest Companies Compare on Diversity?: The

Tech Diversity Scorecard, THE VERGE (Aug. 20, 2015), perma.cc/5XPF-KJJ4. In 2014,

Google began releasing its diversity report to the public along with Facebook, Amazon

and Apple. In 2015, Microsoft released its first diversity report.

. See David McCandless et al., Diversity in Tech: Employee Breakdown of Key

Technology Companies, INFORMATION IS BEAUTIFUL (2017), perma.cc/KHJ2-RUUZ; see also

Visier Insights’ Equal Pay Day Brief Finds Younger Female Workers Lost Ground in 2017,

VISIER (Apr. 10, 2018), perma.cc/92RW-672N.

. Rosie Quinn, Why Traditional Recruitment Methods Are No Longer Enough To

Acquire Top Talent, CIIVSOFT (May 5, 2018), perma.cc/PCB3-ZRC6; Tey Scott, How

Scrapping the Traditional College Recruitment Model Helped LinkedIn Find More Diverse

Talent, LINKEDIN TALENT BLOG (Feb. 6, 2017), perma.cc/E59Q-M5KJ.

. Stephanie Denning, The New Hiring Practices at McKinsey and Goldman Sachs,

FORBES (Apr. 27, 2019), perma.cc/HPC9-5LRC. In addition, 60% of recruiters consider

culture fit one of the most important factors in hiring, which also results in

homogenization. JOBVITE, JOBVITE RECRUITER NATIONAL REPORT 2016: THE ANNUAL SOCIAL

RECRUITING SURVEY 1 (2016), perma.cc/476P-P9A9.

. Luna An et al., Gender Diversity in Tech: Tackling Unconscious Bias, MEDIUM

(Aug. 14, 2017), perma.cc/2795-UEHC (“Unconscious biases are deep-seated ideas and

impressions about certain groups that we carry with us and cause us to draw unfounded

conclusions about people in those groups.”).

. HUM. CAP. INST., INSIGHTFUL HR: INTEGRATING QUALITY DATA FOR BETTER TALENT

DECISIONS (2015), https://perma.cc/EYQ5-W3DV.

. See infra Part II.

. See Anthony G. Greenwald & Linda Hamilton Krieger, Implicit Bias: Scientific

Foundations, 94 CALIF. L. REV. 945, 961 (2006) (“[E]vidence that implicit attitudes

produce discriminatory behavior is already substantial.”).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 294

The responsible use of artificial intelligence,

however, can mitigate

unconscious bias by reducing the impact of human decision-makers on the

process, and create better employment decisions which are based on skills,

traits and behaviors rather than factors (such as sex, race, or pedigree) that

do not correlate with merit or success. A Harris Poll revealed that 75% of

employers reported making a bad hire in the last year.

The responsible use

of artificial intelligence in employment decision-making not only increases

the diversity of candidates and employees, but actually results in more

successful employment outcomes.

AI is the ability of a machine to perform

functions that humans engage in

through the use of a programmed series

of steps known as algorithms. Although there are many domains of AI, as

used herein it refers to algorithms processing data to produce an outcome.

AI can be used to anonymize resumes as well as interviewees, identify

the skills, traits, and behaviors needed to succeed in a certain job, match

applicants with open positions, and predict when an employee is likely to

leave, thereby giving the organization time to remediate the situation and

improve retention.

These measures can attenuate the inherent bias and

. Although others use the terms “people analytics,” “talent management,”

“machine learning,” or “predictive analytics” interchangeably, those terms refer to very

specific processes. “AI” as used herein is intended to reflect the broad category of the use

of computers to perform tasks ranging from removing names from resumes to data

mining performance reviews.

. Ladan Nikravan Hayes, Nearly Three in Four Employers Affected by a Bad Hire,

According to a Recent CareerBuilder Survey, CAREERBUILDER (Dec. 7, 2017),

perma.cc/5BJC-2YUF.

. Charles A. Sullivan, Employing AI, 63 VILL. L REV. 395, (2018). While this paper

focuses on reducing the gender disparity in the tech industry, the author acknowledges

that different issues are encountered by underrepresented minority groups, LGBT+

individuals, and those with disabilities. The author also acknowledges that black women

and others who fall into more than one category face more complex issues around

discrimination than do white women. While this paper is meant to improve conditions

for women overall, further research does need to study the impact of the recommended

methods discussed in this paper on other groups and combinations of groups, but initial

reports confirm that the reduction of unconscious bias and noise in employment

decisions also improves hiring rates for URMs. Guòrun I. Jákupsstova, AI Is Better Than

You at Hiring Diversely, NEXT WEB (May 31, 2018), perma.cc/NQN8-TMFC.

. Lauri Donahue, Commentary, A Primer on Using Artificial Intelligence in the

Legal Profession, HARV. J. L. & TECH. DIG., (Jan. 3, 2018) (“’Artificial Intelligence’ is the term

used to describe how computers can perform tasks normally viewed as requiring human

intelligence, such as recognizing speech and objects, making decisions based on data, and

translating languages.”).

. As an example, Spotify reviews both your previous music selections and the

selections of other users who have chosen similar music in the past. These music

selections are the data, and the recommended songs are the outcome.

. Rohit Punnoose & Pankaj Ajit, Prediction of Employee Turnover in Organizations

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 295

noise present in human decision-making, which are a pervasive problem in

the tech industry.

Additionally, AI can be used to moderate the problem of

human bias baked into the algorithmic process (“algorithmic bias”) by

detecting and correcting problems in biased data sets.

These fixes result

in better accuracy, consistency and fairness in employment decisions.

Most importantly, the use of AI in employment has been shown to increase

the hiring, promotion and retention of women.

As one example, Pymetrics,

which recently received the Technology Pioneer Award from the World

Economic Forum, relies on gamified solutions

which have resulted in a

significant increase in the hiring of women by their clients.

While the term

AI is used throughout, this paper does not suggest that human decision-

making be completely replaced with machines. It proposes that algorithmic-

based decisions are the key to increasing diversity in the tech industry and

Using Machine Learning Algorithms: A Case for Extreme Gradient Boosting, 5 INT’L J.

ADVANCED RES. ARTIFICIAL INTELLIGENCE 22, 26 (2016). For instructions on how to create

your own prediction model, see Marian Dragt, Human Resources Analytics—Predict

Employee Leave, MD2C (Apr. 11, 2018), perma.cc/F2E9-QSRN.

. See infra Part VI.

. See infra Part VII.

. Infor Talent Science, a company that employs AI to collect behavioral

information using a survey, reported a 26% increase in URMs in a sample of 50,000 hires.

Bouree Lam, Recruiters Are Using ‘Algorithmic Hiring’ to Solve One of the Workplaces’

Biggest Problems, BUSINESS INSIDER (June 28, 2015), perma.cc/YF36-UYPL. In an analysis

of seventeen studies comparing human and machine predictions of performance, the

authors concluded that machines were 25% better at evaluating candidates than human,

even when humans had access to more information. Nathan Kuncel et al., Mechanical

Versus Clinical Data Combination in Selection and Admissions Decisions: A Meta-Analysis,

98 J. APPLIED PSYCHOL. 1060, 1060-72 (2013).

. See infra Part VI.

. Gamification is the incorporation of game elements into non-game contexts.

Miriam A. Cherry, The Gamification of Work, 40 HOFSTRA L. REV. 851, 852 (2012).

Gamification in human resources (HR) includes coding challenges, virtual hotel and

restaurant simulations, earning points and badges for completing activities, and a virtual

escape room for assessing collaboration skills. Chiradeep BasuMallick, Gamification in

Recruitment: All You Need to Know, HR TECHNOLOGIST (Nov. 30, 2018), perma.cc/K2P3-

XA5Y; Sara Coene, 9 Examples of Gamification in HR, HR TREND INST. (Feb. 25, 2019),

perma.cc/5PWB-Z5LC.

. Pymetrics Awarded as Technology Pioneer by World Economic Forum, BUSINESS

WIRE (June 21, 2018), perma.cc/EY7J-38ZT. Pymetrics was founded by Harvard and MIT-

trained Ph.Ds, and uses neuroscience to create games which applicants play in order to

be matched with positions. Companies utilizing their services have reported that the

diversity of candidates has increased by 20% and retention by 65%. PYMETRICS,

SUBMISSION TO THE AUSTRALIAN HUMAN RIGHTS COMMISSION: HUMAN RIGHTS AND TECHNOLOGY 2

(Oct. 2, 2018), perma.cc/TMY2-GKL6.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 296

explores solutions for the potential risks noted by various scholars in the

adoption of such programs.

This paper makes three contributions to the intersection of the law,

social science and technology. First, it suggests a way to increase gender

diversity in the tech industry, which is not only the socially responsible

thing to do, but is also the smart thing to do. Second, it provides a solution

to the problem of unconscious bias and noise in human-decision making

without changing or advocating new laws. Third, it explains how AI can

improve employment decision-making and root out and correct

discriminatory outcomes in AI applications. Part I of this paper describes

the environment in the tech industry that women experience and counters

the common reasons expressed for the lack of women in this industry.

Part II explores unconscious bias in gender discrimination law and why it is

an insufficient remedy for the disparities noted in this paper. Part III makes

the business case for women in tech, explaining why it is more than an

equity issue. Part IV examines the failure of current methods to increase

diversity. Part V explains the research on unconscious bias and noise

inherent in human decision-making. Part VI describes why AI is superior to

human decision-making and how it can be implemented to reduce the

impact of unconscious bias and noise. Part VII explains how AI can also be

used to discover and correct the risks of algorithmic bias itself. Part VIII

addresses the legal concerns of using AI in employment decisions, followed

by the conclusion.

II. WOMEN IN THE TECH INDUSTRY

The walk-out of 20,000 Google employees to protest Google’s sexist

culture in November 2018 demonstrates the frustration with the tech

industry’s failure to fulfill its promises.

Although tech companies began

publishing diversity reports in 2014, little has changed, and sexism and

discrimination continue to occur.

A survey of women working in Silicon

. Maria Fernandez, Go Deeper: Google’s Restlessness for Better Company Culture,

AXIOS (Nov. 3, 2018), perma.cc/LG4K-NM5D.

. There has been no meaningful improvement since 2015 when the ELEPHANT IN

THE VALLEY report as referenced infra in note 25 came out. Quentin Fottrell, Woman

Leaders Are Still Getting Screwed by Tech Companies, N.Y. POST (Feb. 8, 2018),

perma.cc/HTK8-8GGM (“Female managers are not only under-represented in

technology companies, they’re also paid significantly less than men. In the Bay Area,

they’re paid $172,585 per year, 10 percent less than men. In Seattle, female managers

are paid an average of $158,858 per year, also 10 percent less than men.”); id.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 297

Valley, The Elephant in the Valley, revealed the lack of diversity and extreme

sexism faced by women at these tech firms, with 88% reporting evidence of

unconscious biases.

The following is a description of some of the problems

women encounter in the tech industry.

A. Issues Faced by Women

One of the issues affecting gender bias in the field is the lack of female

role models and leaders in the industry.

Women make up barely 11% of

tech industry executives

and only 9% of senior IT leadership roles such as

CIOs.

Amazon and Microsoft , headquartered in Washington, reveal a

stunning lack of diversity, especially at senior levels.

Of Amazon’s eighteen

most powerful executives, seventeen are men.

At the recent Consumer

Electronics Show (CES), all of the keynote speakers were male.

especially discouraging fact is that a recent LivePerson survey of 1,000

. Elephant in the Valley, ELEPHANT IN THE VALLEY, perma.cc/97EH-PDB8 (archived

Apr. 19, 2019). Respondents had at least ten years with tech companies and were very

familiar with the myriad gender equity issues. The creators of the survey, from Stanford

and Klein Perkins, wanted to put numbers to the experiences of women in the tech field.

. Kristin Houser, The Tech Industry’s Gender Problem Isn’t Just Hurting Women,

FUTURE SOC’Y (Jan. 31, 2018), perma.cc/94W6-AAKY.

. This lack was noted by 90% of the respondents to a survey by booking.com.

Nick Ismail, Gender Bias in the Tech Industry Is All Encompassing, INFO. AGE (Nov. 8, 2017),

perma.cc/GE8D-6WEB.

. Sheelah Kolhatkar, The Tech Industry’s Gender-Discrimination Problem, NEW

YORKER (Nov. 13, 2017), perma.cc/N5P9-AU78. Women leaders also earn about 10% less

than their male counterparts. Fottrell, supra note 24.

. NAVIGATING UNCERTAINTY: THE HARVEY NASH / KPMG CIO SURVEY 2017 20 (2017),

perma.cc/K3FK-JG7X; see also Luke Graham, Women Take Up Just 9 Percent of Senior IT

Leadership Roles, Survey Finds, CNBC (May 22, 2017), perma.cc/6P3L-7UU3 (finding

virtually no increase in women in IT leadership roles from the previous year in survey of

4,498 CIOs and tech leaders).

. As Bloomberg Businessweek noted, “The search for a second home gives

Amazon something else: an unprecedented opportunity to deal with a problem besetting

all of big tech—a stunning lack of diversity. And Amazon is one of the bigger sinners. Men

make up 73 percent of its professional employees and 78 percent of senior executives

and managers, according to data the company reports to the government. Of the 10

people who report directly to Chief Executive Officer Jeff Bezos, all are white, and only

one—Beth Galetti, the head of human resources—is a woman. The board of directors is

also resisting shareholder pressure to improve gender balance.” Emily Chang, Jeff Green

& Janet Paskin, Amazon Has a Rare Chance to Get More Diverse Fast, BLOOMBERG

BUSINESSWEEK (May 10, 2018), perma.cc/HZ9Q-XJ9P.

. Jason Del Ray, It’s 2017 and Amazon Only Has One Woman Among Its 18 Most

Powerful Executives, RECODE (Oct. 21, 2017), perma.cc/29MQ-DGC6.

. Andrew Mosteller, Female Tech Leaders Take on Equality Issues, LENDIO (Mar. 3,

2018), perma.cc/S7M5-38AU.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 298

people showed that while half of the respondents could name a famous male

tech leader, only 4% could name a female tech leader and one-quarter of

them named Siri and Alexa—who are virtual assistants, not actual people.

The lack of women in leadership roles stems from the inability for most

women to move up in these companies. This glass ceiling, as well as

widespread disrespect, harassment and exclusion, results in about half of

the women

entering the field to leave it (compared with 17% of men).

The Elephant in the Valley survey indicated that 87% of the women reported

receiving demeaning comments from male colleagues, 47% said they had

been asked to do lower-level tasks that male colleagues were not asked to

do, and 66% said they had been excluded from important social or

networking events.

Comments on the survey indicated that women were

disrespected in numerous ways, such as being asked to take notes at

meetings or order food, and being ignored in favor of male subordinates

during meetings.

While this figure does not surprise most women, men seem

flabbergasted

that 90% of the women surveyed reported witnessing sexist

. Monica Torres, Survey Reveals People Think Siri and Alexa Are Female Tech

Leaders, THE LADDERS (Mar. 23, 2018), perma.cc/4TDM-BB5C.

. Hickey, supra note 2; see also Glass et al., supra note 2.

. Houser, supra note 2625.

. Id. For some of the more outrageous experiences, see EMILY CHANG, BROTOPIA:

BREAKING UP THE BOY’S CLUB OF SILICON VALLEY (2018).

. Elephant in the Valley, supra note 25. In another interesting experiment, two

co-workers, one male and one female, exchanged email signatures to see if they were

being treated differently because of their sex. The male using the female email signature

said the treatment was like night and day. As “Nicole,” he found clients to be rude,

dismissive, and condescending. Nicole indicated that she had the most productive week

of her life because she did not have to convince clients to respect her. Nicole Hallberg,

Working While Female, MEDIUM (Mar. 9, 2017), perma.cc/2C3M-NPYW; see also

@lindworm, Working as a Woman Can #Suck, TWITTER (Mar. 9, 2017), perma.cc/27TK-

4WV8 (displaying actual string of tweets). This sexism extends to women in positions of

leadership as well. Jenny Campbell, who sold her business for £50 million, was

repeatedly misidentified as the wife of her male subordinate when he was standing next

to her. Jamie Johnson, I Sold My Business for £50 Million, But It's Still Assumed I Am the

Wife of the Man Standing Next to Me in Meetings, TELEGRAPH (Mar. 23, 2018),

perma.cc/A4ZF-M6XP. In order to avoid this exact problem, Penelope Gazin and Kate

Dwyer, the owners of a successful online marketplace, invented an imaginary co-founder

named Keith. While some of the obstacles they faced were overt, such as the web

developer who deleted their site after one of the owners indicated a lack of interest in

dating him, most of their experiences were the result of subtle sexism. While it took days

for the women to receive input from collaborators, “Keith” received responses

immediately. John Paul Titlow, These Women Entrepreneurs Created A Fake Male

Cofounder to Dodge Startup Sexism, FAST COMPANY (Aug. 29, 2017), perma.cc/HE9Z-T6V3.

. Alice Berg, Gender Discrimination Problem: Silicon Valley VS Women,

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 299

behavior at their company and industry conferences.

Sixty percent had

been harassed themselves and 33% feared for their safety due to work-

related circumstances.

However, most of these incidents are not reported

due to fear of retaliation.

In one well-publicized case, AJ Vandermeyden,

who sued Tesla for discrimination, retaliation, and other workplace

violations, was fired after filing suit.

Therese Lawless, Vandermeyden’s

attorney, confirmed that firing whistleblowers is a common form of

retaliation against women who complain of discrimination. “That’s the

message the company sends if you speak up. That’s why people are fearful,”

Lawless said.

In addition to sexism, there is also the issue of stereotyping.

Characteristics that tend to be valued in men, often resulting in the

advancement in their careers, have the opposite effect when exhibited by

women. Eighty-four percent of those surveyed reported that they had been

told they were “too aggressive,” and 66% reported being excluded from key

social/networking opportunities because of their gender.

While peers

prefer successful men to unsuccessful ones, successful women are

MUNPLANET (Jan. 10, 2018), perma.cc/Y98L-4WPS.

. Elephant in the Valley, supra note 25.

. Houser, supra note 26.

. Kolhatkar, supra note 28. Most companies, upon hearing a report from a

woman, will ask her to “prove” the event occurred if the man denies it happened. CHAI R.

FELDBLUM & VICTORIA A. LIPNIC, U.S. EQUAL EMP’T OPPORTUNITY COMM’N, REPORT OF THE CO-

CHAIRS OF THE EEOC SELECT TASK FORCE ON THE STUDY OF HARASSMENT IN THE WORKPLACE (2016).

Because there is no audio or video proof, nor witnesses, no action is taken and the

woman eventually leaves the company because of being shunned, demoted, or

transferred after making the report, or chooses to leave because of the firm’s disrespect

shown by choosing to believe the perpetrator rather than the woman harassed.

According to Melinda Gates, the co-chair of the Bill & Melinda Gates Foundation, “Men

who demean, degrade or disrespect women have been able to operate with such

impunity—not just in Hollywood, but in tech, venture capital, and other spaces where

their influence and investment can make or break a career[.] The asymmetry of power is

ripe for abuse.” Kolhatkar, supra note 28.

. Kolhatkar, supra note 28. Ellen Pao, whose discrimination case was widely

publicized, was one of the women who found herself the target of harassment by a male

colleague she briefly dated. After filing suit, Pao was terminated and her complaint was

amended to include a count of retaliation. Although Pao lost her case in 2015, everyone

got a glimpse of the vitriolic response of the venture capital firm she worked for, Kleiner

Perkins, which only confirmed in the public’s eye the likelihood that Pao was accurately

describing what she had encountered. Eric Johnson, Why Did Ellen Pao Lose Her Gender

Discrimination Lawsuit? ‘People Were Not Ready’, VOX (Oct. 2, 2017),

https://perma.cc/YS4T-QAM4.

. Id.

. Elephant in the Valley, supra note 25.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 300

penalized.

In employee reviews, women are referred to as “abrasive,”

“aggressive,” “bossy,” and “strident.”

Women who attempt to negotiate

salaries are viewed as being “difficult to work with” even when using the

same language as men.

This is known as the likeability penalty.

One of the most interesting findings I came across was that many men

do not acknowledge that a gender diversity problem even exists.

If men,

who hold 80% of the leadership positions in tech companies, do not even

believe that the low levels of women in their company is a problem, it is

unlikely it will get resolved. According to a 2017 study, 50% of men reported

that it is sufficient when just one in ten senior leaders in their company is a

woman.

What is especially telling is a study showing when 17% of the

people in a room are women, men report that they think the room is 50-50

men and women.

When 33% of the people in the room are women, men

believe they are outnumbered.

This skewed view of reality may explain

why men in tech express a belief that there are enough women in leadership

positions in their companies already

despite women only comprising 11%

of executives in tech as indicated above.

. Successful women are viewed as being less likeable. While terms like

“confident” and “strong” are used to describe successful men, women are called “bossy”

and “aggressive.” 7 Tips for Men Who Want to Support Equality, LEANIN.ORG,

perma.cc/7KMG-5M67. In tech, these types of terms appear in 85% of female high

performers’ evaluations compared to only 2% of the men’s evaluations. Kieran Snyder,

The Abrasiveness Trap: High-Achieving Men and Women Are Described Differently in

Reviews, FORTUNE (Aug. 26, 2014) [hereinafter Snyder, Abrasiveness], perma.cc/VB7V-

LX7W.

. Snyder, Abrasiveness, supra note 45.

. Hannah Riley Bowles et al., It Depends Who is Asking and Who You Ask: Social

Incentives for Sex Differences in the Propensity to Initiate Negotiation—Sometimes It Does

Hurt to Ask, 103 ORGANIZATIONAL BEHAV. & HUM. DECISION PROCESSES 84 (2007).

. Marianne Cooper, For Women Leaders, Likability and Success Hardly Go Hand-

in-Hand, HARV. BUS. REV. (Apr. 30, 2013), perma.cc/MBD9-HHPK.

. Lauren Williams, Facebook’s Gender Bias Goes So Deep It’s in the Code, THINK

PROGRESS (May 2, 2017), perma.cc/9M9Z-GZJH.

. ALEXIS KRIVKOVICH ET AL., MCKINSEY & CO., WOMEN IN THE WORKPLACE 2017

(Oct. 2017), perma.cc/J2TG-UD2R.

. Linnea Dunne, So You Think You Were Hired on Merit? Gender Quotas and the

Perception Gap, LINNEA DUNNE BLOG (Aug. 21, 2017), perma.cc/WLW3-LUKS.

. Id.

. Emma Hinchliffe, 58% of Men in Tech Say There Are Enough Women in

Leadership Roles, But Women Don’t Agree, MASHABLE (Sept. 20, 2017),

https://perma.cc/3BPG-2MWN. When a study was shown to male faculty members

demonstrating the unjustified preference for male lab managers (where simply changing

names from female to male on a resume made the lab manager more likely to be hired),

they still assessed bias against women as being low. Alison Coil, Why Men Don’t Believe

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 301

B. Reasons Attributed to the Lack of Women in the Tech Industry

Although women hold 57% of professional occupations in the U.S.

workforce, they occupy only 26% of professional computing jobs.

Reasons

alleged to explain this gender gap in the tech field include: lack of pipeline,

lack of interest, and the confident assertion of meritocracy.

Although

women with STEM degrees are available,

companies hire men with

science and engineering degrees at twice the rate of women.

One study

shows that when a tech job was available, 53% of the time, companies

interviewed no female candidates at all.

Women also receive lower

salaries for the same job at the same company 60% of the time.

Although women are interested in these jobs, many are alienated during

the recruiting process itself.

Researchers from Stanford revealed that

The Data On Gender Bias In Science, WIRED (Aug. 25, 2017), https://perma.cc/RNQ6-

PWSX. To review the actual study, see Corinne A. Moss-Racusin et al., Science Faculty’s

Subtle Gender Biases Favor Male Students, 109 PROCS. NAT’L ACAD. SCI. 16,474 (2012).

. NAT’L CENTER FOR WOMEN & INFO. TECH., BY THE NUMBERS (2019),

https://perma.cc/J4Y5-7B7R.

. Banu Ozkazanc-Pan, Women in Tech Suffer Because of the American Myth of

Meritocracy, THE LADDERS (May 4, 2018), https://perma.cc/XK4F-5PZC; Reshma Saujani

& Ayah Bdeir, Opinion: You Can't Solve the Gender Gap in Tech If You Don't Understand

Why It Exists, BUZZFEED NEWS (Mar. 6, 2019), https://perma.cc/84BT-VD8C.

. At Cornell, 55% of the incoming freshmen in the fall of 2018 in the engineering

school who indicated an interest in computer science were women. The year before, 38%

of declared computer science majors were women. 55 Percent of Incoming Eng Students

Interested in Computer Science Are Women, CORNELL CIS, https://perma.cc/BA7Z-DULH

(archived Apr. 13, 2019) [hereinafter 55 Percent]. Dartmouth College graduates more

women in computer science than men, at 54%. Thea Oliver, An In-Depth Look at the

Gender Gap in the Tech Industry, TECHNICALLY COMPATIBLE (May 12, 2017),

https://perma.cc/E7G4-XR6N. Harvey Mudd graduated 56% women in computer

science in 2018. Harvey Mudd Graduates Highest-Ever Percentage of Women Physics and

Computer Science Majors, HARVEY MUDD COLL. NEWS (May 15, 2018),

https://perma.cc/Y5TW-QX9Z; see also Kristen V. Brown, TechShift: More Women in

Computer Science Classes, SFGATE (Feb. 18, 2014), https://perma.cc/P2EN-24KJ

(“Berkeley, Stanford and a handful of other universities have experienced a marked

uptick in the numbers of female computer science students.”). In addition, outside of the

United States, women in STEM are the rule, not the exception. In Iran, 70% of STEM

graduates are female, with over 60% in Oman, Saudi Arabia, and the UAE. Annalisa

Merelli, The West Is Way Behind Iran and Saudi Arabia When It Comes to Women in

Science, QUARTZ (Mar. 8, 2018), https://perma.cc/QD5Z-WY8N.

. Liana Christin Landivar, Disparities in STEM Employment by Sex, Race, and

Hispanic Origin, AM. COMM. SURV. REP. (Sept. 2013), https://perma.cc/TB4M-RQ62.

. NANCY M. CARTER & CHRISTINE SILVA, CATALYST, THE MYTH OF THE IDEAL WORKER: DOES

DOING ALL THE RIGHT THINGS REALLY GET WOMEN AHEAD? (2011).

. The State of Wage Inequality in the Workplace, HIRED (2019),

https://perma.cc/HS75-8E6C.

. Alison T. Wynn & Shelley J. Correll, Puncturing the Pipeline: Do Technology

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 302

“through gender-imbalanced presenter roles, geek culture references, overt

use of gender stereotypes and other gendered speech and actions,

representatives may puncture the pipeline, deflating the interest of women

at the point of recruitment into technology careers.”

In addition, the

wording of job ads can dissuade women from applying. Researchers

discovered that women were less likely to apply to jobs described with

masculine words such as “competitive” and “dominant.”

Gendered job ads

resulted in women believing 1) the company employed more men than

women, 2) they did not belong, and 3) the job would not be appealing.

Even when women are hired, the tech companies are unable to retain

them.

Women leave the tech field at a rate 45% higher than men do.

the very low levels of women who are hired, half leave because of the work

environment.

A longitudinal study regarding retention reveals the real

problem, and it is not the pipeline.

In a survey of 716 women in tech who

left the field, all of them said they enjoyed the work, but not the workplace

environment.

Women increasingly are speaking out about what they see

as a hostile culture due to buddy networks.

These informal networks,

which benefit men, tend to exclude otherwise qualified women.

Companies Alienate Women in Recruiting Sessions? 48 SOC. STUD. SCI. 149 (2018),

https://perma.cc/JK3B-B4JD.

. For an explanation of how recruiters from tech companies alienate female

applicants, see Wynn et al., supra note 60 (84% of presenters were male, women

representatives were used to hand out swag, presenters promoted the fraternity-like

environment, and in one case, a presenter made multiple references to pornography and

prostitution).

. Danielle Gaucher et al., Evidence That Gendered Wording in Job Advertisements

Exists and Sustains Gender Inequality, 101.1 J. PERSONALITY & SOC. PSYCHOL. 109 (2011).

. Id.

. Kim Williams, Women in Tech: How to Attract and Retain Top Talent, INDEED BLOG

(Nov. 6, 2018), perma.cc/7AY8-D8ZU.

. Williams, supra note 64.

. Kieran Snyder, Why Women Leave Tech: It's the Culture, Not Because ‘Math Is

Hard’, FORTUNE (Oct. 2, 2014), perma.cc/F5XF-KZXC [hereinafter Snyder, Women Leave

Tech].

. Glass, supra note 2.

. Snyder, Women Leave Tech, supra note 66. See also Kolhatkar, supra note 28, for

a discussion of what women experience in the workplace environment.

. Chang, supra note 36.

. Laura Colby, Women and Tech, BLOOMBERG: QUICKTAKE (Aug. 8, 2017),

https://perma.cc/D974-DTSM.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 303

“Undermining behavior from managers” was also reported as a major

factor,

as well as the inability to move up in the company.

While women initially are attracted to the tech industry, structural

barriers to advancement and workplace issues force them out.

A 2018

report from McKinsey found that women made up 48% of entry-level roles

in tech with only 23% of those advancing to senior management roles.

Men are not only promoted more frequently, even when a woman is given

credit for her contributions to the growth of the company, she does not

receive the promotion.

Men also tend to receive recognition more often

than women.

While men are hired based on their potential, women are

hired based on their proven experience.

This is known as the “prove it

again syndrome.”

. Liza Mundy, Why Is Silicon Valley So Awful to Women?, THE ATLANTIC, Apr. 2017.

. Tekla S. Perry, Women Leave Tech Jobs Because They Can’t Climb the Ladder,

IEEE SPECTRUM (Nov. 6, 2018), https://perma.cc/LFS3-8M4L.

. See Allison Schnidman, Why Women Are Leaving Their Jobs (Your First Guess Is

Wrong), LINKEDIN TALENT BLOG (Nov. 5, 2015), https://perma.cc/3XW9-NE2F (revealing

that, according to a survey of 4,000 women in the tech industry, the top three reason

women left their tech jobs were concern for the lack of opportunities for advancement,

dissatisfaction with leadership and the work environment). See also CAROLINE SIMARD ET

AL., ANITA BORG INST. FOR WOMEN & TECH., CLIMBING THE TECHNICAL LADDER: OBSTACLES AND

SOLUTIONS FOR MID-LEVEL WOMEN IN TECHNOLOGY (2008) (providing results from a survey of

female mid-level managers at Silicon Valley high-tech firms regarding barriers to

advancement).

. ALEXIS KRIVKOVICH ET AL., MCKINSEY & CO., WOMEN IN THE WORKPLACE 2018 (2018).

. Berg, supra note 38.

. According to a survey of 1,000 professionals, 50% of men indicated that they

received recognition at work at least a few times per month compared with only 43% of

women. Bryson Kearly, Is Gender Equality in the Workplace Still an Issue? Studies Say Yes!,

HR INSIGHTS (Apr. 20, 2016), https://perma.cc/K3EA-XLRW.

. Joan C. William et. al., Tools for Change: Boosting the Retention of Women in the

STEM Pipeline, 6 J. RES. GENDER STUD. 11 (2016).

. Eileen Pollack covered this in detail in a New York Times article describing the

“Prove-It-Again! bias” which requires women to provide more evidence of competence

than men in order to be seen as equally competent. Pollack points out a study conducted

at Yale proving that a young male scientist will be viewed more favorably than a woman

with the same qualifications. When professors at six major research institutions were

presented with identical summaries of the accomplishments of two imaginary

applicants, they were significantly more willing to offer the man a job. When they did

choose a woman, they paid her $4,000 less than what they paid the men hired. In keeping

with Pollack’s findings, a peer-reviewed study of top U.S. graduate programs in the

sciences funded by the National Academy of Sciences demonstrated that both female and

male professors rated male applicants for a lab manager position as “significantly more

competent and hirable than (identical) female applicants.” Eileen Pollack, Why Are There

Still So Few Women in Science?, N.Y. TIMES MAG. (Oct. 3, 2013), https://perma.cc/6FRA-

74VB.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 304

While few come right out and say that women lack ability, the

explanation most often used to disguise the prejudice against women is that

the tech industry is a “meritocracy” implying that the men are simply more

qualified.

This argument does not stand up to scrutiny. A study of 1.6

million students showed that the top 10% of STEM classes contain an equal

number of men and women.

In terms of actual skills related to the position,

women may in fact be better at coding than men. A study reflecting data

from the largest open source community (GitHub) with 12 million

collaborators across 31 million software repositories showed that while

women’s codes were rated more harshly than men’s when gender was

visible, when gender was hidden, the women’s codes were found to be rated

consistently better.

This study refutes the argument that women are

somehow less qualified or capable than men and demonstrates how the

meritocracy argument is largely a reflection of gender bias rather than

actual verifiable fact.

Because decision-makers are unaware of their own biases, they explain

their decision as being “on the merits” without factoring in their preference

for a candidate based on factors that have nothing to do with job skills.

addition, decision-makers may focus their attention on information that

confirms their existing belief system and disregard potentially relevant

information that would tend to contradict it.

“Most interviews are a waste

. In a now-deleted article on Forbes, tech writer Brian S. Hall argued that Silicon

Valley was in fact a meritocracy. He stated, “If you aren’t able to make it here, it’s almost

certainly not because of any bias,” and argued anyone claiming bias should blame their

own “refusal to put in the hard work.” Dexter Thomas, Forbes Deleted a White Tech

Writer’s Article That Called Silicon Valley a ‘Meritocracy’, L.A. TIMES (Oct. 8, 2015),

https://perma.cc/E74T-3HNQ. See also Is Tech a Meritocracy?, https://perma.cc/F23A-

RQ3A (archived May 8, 2019) (providing numerous criticisms of the allegation that tech

is a meritocracy).

. R. E. O’Dea et al., Gender Differences in Individual Variation in Academic Grades

Fail to Fit Expected Patterns for STEM, 9 NATURE COMM. 3777 (2018).

. Josh Terrell et al., Gender Differences and Bias in Open Source: Pull Request

Acceptance of Women Versus Men, 3 PEER J. COMP. SCI. e111 (2017). See also Julia Carriew,

Women Considered Better Coders—But Only If They Hide Their Gender, THE GUARDIAN

(Feb. 12, 2016), https://perma.cc/KEN7-VZNT (describing GitHub research study).

. For example, someone who graduated from Harvard may exhibit a preference

for a candidate who also attended Harvard. Julia Mendez, The Impact of Biases and How

to Prevent Their Interference in the Workplace, INSIGHT INTO DIVERSITY (Apr. 27, 2017),

https://perma.cc/JAD3-9R4J.

. Kathleen Nalty, Strategies for Confronting Unconscious Bias, 45 COLO. LAW. 45

(2016). “Another type of unconscious cognitive bias—attribution bias—causes people to

make more favorable assessments of behaviors and circumstances for those in their ‘in

groups’ (by giving second chances and the benefit of the doubt) and to judge people in

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 305

of time because 99.4% of the time is spent trying to confirm whatever

impression the interviewer formed in the first 10 seconds,” according to

Laszlo Bock, the author of Work Rules!

Similarly, companies who tout

meritocracy actually demonstrate more bias against women than those who

do not.

The solution discussed in Part VII below is the use of AI to

moderate this type of human error from the hiring process.

Unconscious biases and noise not only influence employment decisions,

but also how the workplace culture evolves.

The effect of unconscious

biases is well correlated with discriminatory employment decisions.

Although studies bear this out, the courts have a difficult time reconciling

these subtler forms of discrimination with the law.

III.GENDER DISCRIMINATION LAW

Title VII of the Civil Rights Act prohibits the discrimination by an

employer against any individual with respect to compensation, terms,

conditions, or privileges of employment, because of such individual’s race,

color, religion, sex, or national origin.

Although overt forms of

discrimination have been reduced due to antidiscrimination law and

changes in societal norms,

cases involving more covert forms of

their ‘out groups’ by less favorable group stereotypes.” Id. “The adverse effects of many

of these cognitive biases can be compounded by affinity bias, which is the tendency to

gravitate toward and develop relationships with people who are more like ourselves and

share similar interests and backgrounds. This leads people to invest more energy and

resources in those who are in their affinity group while unintentionally leaving others

out.” Id.

. Jennifer Alsever How AI Is Changing Your Job Hunt, FORTUNE (May 19, 2017),

https://perma.cc/24FG-T8G2.

. Emilio J. Castilla & Stephen Benard, The Paradox of Meritocracy in Organizations,

55 ADMIN. SCI. Q. 543 (2010). A study at Cornell revealed that when the participants were

asked to award bonuses to men and women with similar profiles, telling them that their

company valued merit-based decisions actually increased the likelihood of higher

bonuses to the men. Id.

. An, supra note 7.

. See Melissa Hart, Subjective Decisionmaking and Unconscious Discrimination, 56

ALA. L. REV. 741, 744-745 (2004) (“There is little doubt that unconscious discrimination

plays a significant role in decisions about hiring, promoting, firing, and other benefits

and tribulations of the workplace”); Audrey Lee, Unconscious Bias Theory in Employment

Discrimination Litigation, 40 HARV. C.R.-C.L. L. REV. 481, 483-87 (2005) (“Courts have

recognized the existence of unconscious discrimination since the earliest Title VII

decisions”); Greenwald & Krieger, supra note 10.

. 42 U.S.C. § 2000e-2(a) (2012).

. Lee, supra note 87at 488.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 306

discrimination have been less successful. As such, current application of law

does not provide an adequate remedy for those harmed by non-obvious

non-intentional discrimination.

Class action suits for these types of

matters are seldom certified,

and most tech companies have arbitration or

confidentiality requirements that prevent women from getting their day in

court.

Although social science has greatly advanced our understanding of how

unconscious biases influence the workplace and can lead to

discrimination,

courts have been inconsistent in their treatment of this

evidence.

Because courts have required proof of “intent” in disparate

treatment cases,

most actions relying on unconscious bias as the cause of

an adverse action, assert a disparate impact claim.

However, cases relying

. Intentional discrimination theory would not cover harms due to subjective

human decision-making because “intent” requires some outward showing of prejudice

resulting in a protected group being subjected to an adverse employment action due to

their membership in the group. Stephanie Bornstein, Reckless Discrimination, 105 CALIF.

L. REV. 1055 (2017).

. See discussion in Subpart II.A.

. IMRE S. SZALAI, THE EMP. RIGHTS ADVOC. INST. FOR LAW & POL’Y, THE WIDESPREAD USE OF

WORKPLACE ARBITRATION AMONG AMERICA’S TOP 100 COMPANIES 6 (2018).

. Linda Hamilton Krieger, The Content of Our Categories: A Cognitive Bias

Approach to Discrimination and Equal Employment Opportunity, 47 STAN. L. REV. 1161

(1995) (suggesting that biased discriminatory decisions may result not from intentional

action, but rather “unintentional categorization-related judgment errors characterizing

normal human cognitive function.”).

. See Anthony Kakoyannis, Assessing the Viability of Implicit Bias Evidence in

Discrimination Cases: An Analysis of the Most Significant Federal Cases, 69 FLA. L. REV. 1181

(2018) (discussing the differences in the treatment of implicit bias evidence by the

courts); Melissa Hart and Paul M. Secunda, A Matter of Context: Social Framework

Evidence in Employment Discrimination Class Actions, 78 FORDHAM L. REV. 37, 50 (2009)

(providing cites to courts which have permitted certification and those which have not).

. Disparate treatment occurs when employees can show that they were treated

differently than those who are not members of the same protected class. To assert this

cause of action, courts require that plaintiffs show that their employer engaged in

“intentional” discrimination by taking an adverse employment action on the basis of

membership in the protected class. However, the defendant employer is able to avoid

liability by demonstrating a nondiscriminatory justification for the action. The plaintiff

employees would still be able to prevail if they can show that the justification was simply

a pretext. Bornstein, supra note 90.

. In 1971, the Supreme Court in Griggs v. Duke Power Co., 401 U.S. 424 (1971)

first enunciated the disparate impact theory of discrimination. Under disparate impact

theory, an employment practice that is neutral on its face, but in application has a

disproportionately negative effect on a statutorily protected group is unlawful, unless

the employer can prove that the practice is job-related and a business necessity. Id. at

431. However, liability can still attach if the plaintiff can show an alternative less

discriminatory practice. Wards Cove Packing Co. v. Atonio, 490 U.S. 642, 644 (1989). See

Stewart v. City of St. Louis, 2007 U.S. Dist. LEXIS 38473, at *22 n.4 (E.D. Mo. 2007); 42

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 307

on unconscious bias evidence to certify class action lawsuits have not been

uniformly successful due to inconsistencies in how lower courts have

interpreted Wal-Mart v. Dukes.

A. Unconscious Bias in Case Law

In Wal-Mart v. Dukes, some 1.5 million female Wal-Mart employees

alleged in a class action complaint that the company discriminated against

them by denying women equal pay and promotions.

Wal-Mart did not

have any testing procedures in place for evaluating employees and used

discretionary local decision-making with respect to employment matters.

The plaintiffs alleged this store-level discretion violated Title VII. The

Supreme Court refused to allow the certification of the class explaining that

there was “no common question” among the 1.5 million plaintiffs

despite

social science evidence explaining how local subjective decision-making

resulted in the lower pay and lack of promotions of its female employees

due to the unconscious biases of the decision-makers.

100

After the Wal-Mart case, it was uncertain whether unconscious bias

evidence would be allowed with respect to class action certification relying

on statistical analysis.

101

However, the court in Ellis v. Costco Wholesale

USCS § 2000e-2(k) (2012). With a disparate impact case, the plaintiff is not required to

show intent, but rather the impact of the employment decision can prove discrimination.

Griggs, 401 U.S. at 432. This standard is codified at 42 U.S.C.S. § 2000e-2(k) (2012).

. 564 U.S. 338 (2011). See Annika L. Jones, Implicit Bias as a Social-Framework

Evidence in Employment Discrimination, 165 U. PA. L. REV. 1221, 1231 (2017)

(discussing inconsistent treatment of the use of social science evidence in support of the

commonality requirement in class action certification after Wal-Mart).

. 564 U.S. at 338.

. Rules 23(a) and (b) of the Federal Rules of Civil Procedure set forth the

requirements for class certification. Rule 23(a) requires: (1) the class is so numerous

that joinder of class members is impracticable; (2) there are questions of law or fact

common to the class; (3) the claims or defenses of the class representatives are typical

of those of the class; and (4) the class representatives will fairly and adequately protect

the interests of the class.

100

. See Camille A. Olson et al., Implicit Bias Theory in Employment Litigation, 63

PRAC. LAW. 37 (2017) (explaining the Wal-Mart decision and contrasting it with cases

where implicit bias theory was accepted).

101

. See Tanya K. Hernandez, One Path for Post-Racial Employment Discrimination

Cases—The Implicit Association Test Research as Social Framework Evidence, 32 LAW &

INEQ. 309, 332-35 (2014) (noting that the failure of courts to recognize implicit bias

evidence leaves plaintiffs without a remedy for discrimination due to unconscious

thoughts or feelings that influence their decisions).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 308

Corp.

102

granted the employees’ motion for class certification based in part

upon unconscious bias expert testimony that the employer’s corporate

culture created and reinforced stereotyped thinking, which allowed gender

bias to infect the promotion process from leadership down. Although in

Wal-Mart v. Dukes, the court concluded that the plaintiffs were unable to

show that the statistical analysis evidence of unconscious bias related

specifically to Wal-Mart’s employment practices, and thus was insufficient

to prove the existence of questions of law or fact common to the particular

proposed class per Fed. R. Civ. P. 23(a),

103

the court did find it in Ellis at least

for class certification purposes.

104

Costco ultimately settled the case for $8

million after the case was remanded.

105

B. Unconscious Bias in Cases That Were Settled

Although case law has been inconsistent in the treatment of

unconscious bias testimony in the court room since Wal-Mart, such claims

have had enough traction to result in significant out-of-court settlements.

The Court’s receptivity to unconscious bias arguments in the Home Depot

and FedEx class action suits resulted in those cases settling for considerable

amounts, $87.5 million and $53.5 million respectively.

106

Unconscious bias

has been raised in a number of other class actions against Fortune 500

102

. 285 F.R.D. 492 (N.D. Cal. 2012).

103

. Wal-Mart, 564 U.S. at 356; Christine A. Amalfe, The Limitations of Implicit Bias

Testimony—Post Dukes, GIBBONS P.C. (Mar. 2013), https://perma.cc/5PH3-38LF.

104

. Order Granting Plaintiff’s Motion for Class Certification; and Denying

Defendant’s Motion to Eliminate Class Claims, Ellis v. Costco Wholesale Corp., 285 F.R.D.

492, (N.D. Cal. 2012) (No. C-04-3341 EMC).

105

. Order Granting Motion for (1) Preliminary Approval of Class Action Settlement;

(2) Approval of Class Notice and Notice Plan; and (3) Setting of Schedule for Final

Approval at Exhibit 1 § 3; Ellis v. Costco Wholesale Corp., 285 F.R.D. 492 (N.D. Cal. 2012)

(No. C04-3341 EMC).

106

. Dana Wilkie, Tips for Rooting Out Hidden Bias, SOC’Y FOR HUM. RES. MGMT (Dec. 1,

2014), https://perma.cc/DD8M-JVEP.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 309

companies such as American Express,

107

Morgan Stanley,

108

General

Electric,

109

Best Buy,

110

Bank of America,

111

and Cargill.

112

The EEOC has also taken a special interest in the tech industry

113

and in

rooting out unconscious bias in the workplace. “As the EEOC looks ahead to

the next 50 years, there are two main issues to solve. First is helping

companies create a more robust talent pipeline for women and minorities

with greater representation at every level of management. This includes

identifying and eliminating unconscious and conscious bias in the

workplace. This could also include more advanced analytics to assess

systematic discrimination and patterns of practice both at the company

and the industry level. [emphasis added].”

114

Due to increased awareness

of unconscious bias and the failure of the tech industry to meaningfully

107

. In 2002, American Express Financial Advisors agreed to pay $31 million to

4,000 female financial advisers who had filed a class action suit for discrimination

against them. Jonathan D. Glater, American Express Unit Agrees to Settle Sex

Discrimination Case, N.Y. TIMES (Feb. 21, 2002), https://perma.cc/H284-3JYB.

108

. In 2004, Morgan Stanley agreed to pay $54 million to 340 women to settle a sex

discrimination case rather than stand trial on the EEOC’s suit that alleged it denied equal

pay and promotions to women in a division of its investment bank. Dan Ackerman,

Morgan Stanley: Big Bucks for Bias, FORBES (Jul. 13, 2004), https://perma.cc/3AWE-

WHPG.

109

. In 2009, a former high-ranking attorney for General Electric Co. settled her class

action gender discrimination lawsuit on behalf of 1,500-1,700 female employees alleging

system company-wide gender discrimination against the company for an undisclosed

amount. John Christoffersen, Former GC Settles Gender Discrimination Suit Against

General Electric, LAW.COM (Jan. 28, 2009), https://perma.cc/KNL9-YAM5.

110

. In 2011, Best Buy agreed to settle a class action lawsuit accusing the largest U.S.

electronics retailer of job discrimination, paying a total of $200,000 to the nine named

plaintiffs plus as much as $10 million for legal fees and costs. REUTERS, Best Buy Settles

Class-Action Racial Job Discrimination Lawsuit, HUFFPOST (June 6, 2011),

https://perma.cc/F57U-TMXX.

111

. In 2013, Bank of America agreed to pay $39 million to 4,800 women who

worked in its Merrill Lynch brokerage operation. Patrick McGeehan, Bank of America to

Pay $39 Million in Gender Bias Case, N.Y. TIMES (Sept. 6, 2013), https://perma.cc/7QCG-

2WUG.

112

. In 2014, Cargill Meat Solutions agreed to pay $2.2 million to settle

discrimination charges brought by the Department of Labor’s Office of Federal Contract

Compliance Programs (OFCCP). Daniel McCoy, Cargill to Pay $2.2 Million to Settle

Discrimination Claims, WICHITA BUS. J. (Jan. 23, 2014), https://perma.cc/A4HS-J53A.

113

. U.S. GOV’T ACCOUNTABILITY OFFICE, DIVERSITY IN THE TECHNOLOGY SECTOR: FEDERAL

AGENCIES COULD IMPROVE OVERSIGHT OF EQUAL EMPLOYMENT OPPORTUNITY REQUIREMENTS

(2017).

114

. Meeting of May 18, 2016—Promoting Diverse and Inclusive Workplaces in the

Tech Sector, U.S. EQUAL EMP. OPPORTUNITY COMMISSION (May 18, 2016),

https://perma.cc/2P6W-M3N4.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 310

increase diversity, these organizations should be very concerned with the

potential for unconscious bias discrimination suits.

115

IV.THE BUSINESS CASE FOR WOMEN IN TECH

There are a number of reasons why the tech industry should be

disturbed with the lack of women and under-represented minorities

(URMs) in their ranks. In addition to potential lawsuits, EEOC investigations,

and the inability to hire all of the employees they will need in the near

future, there is also great value in creating a diverse workforce. Studies

show that women in leadership roles produce significant benefits to

companies, including better decision-making, improved working

environment, a more collegial atmosphere, and increased innovation.

116

is well-established that “diverse teams outperform homogeneous teams.”

117

Decision-making research shows that diverse teams can avoid group-

think.

118

Women are better able to examine different points of view and

consider different perspectives than men are.

119

It also does not make sense

fiscally to treat women in this way. Research consistently shows that

companies with high gender diversity are more profitable and less volatile

those with low gender diversity.

120

Based on the available statistics about

the benefits of women in tech, it seems counterintuitive for the tech industry

to have such a gender imbalance.

121

Tech companies seem to be on the

cutting edge, anticipating what the public wants even before the public

115

. In 2015, Katie Moussouris initiated a class action suit against Microsoft alleging

gender discrimination which is still pending. The complaint noted that she and other

women earned less than their male counterparts and that men were given preferential

treatment in promotions resulting of the unconscious biases of decision-makers.

Plaintiff’s Motion for Class Certification at 1, Moussouris v. Microsoft Corp., 311 F. Supp.

3d 1223 (W.D. Wa. 2018) (No. C15-1483JLR).

116

. Why It Pays to Invest in Gender Diversity, MORGAN STANLEY (May 11, 2016),

https://perma.cc/5CCA-XKQL.

117

. Katherine Phillips, How Diversity Makes Us Smarter, SCI. AM. (Oct. 1, 2014),

https://perma.cc/U4VZ-XXA7.

118

. Anna Johansson, Why Workplace Diversity Diminishes Groupthink and How

Millennials Are Helping, FORBES (Jul. 20, 2017), https://perma.cc/RQ4J-PWUP.

119

. Nancy Wang, Diversity Is the Key to Startup Success—What Can Early-Stage

Founders Do About It?, FORBES (Nov. 12, 2018), https://perma.cc/DK3X-T3J4.

120

. Phillips, supra note 117.

121

. See Erin Griffith, There’s a Simple Way to Improve Tech’s Gender Imbalance,

FORTUNE (June 1, 2016), https://perma.cc/4254-T3CY (stating that women influence

household spending and are more likely than men to adopt technology trends early).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 311

knows it.

122

Given the reputation loss the tech industry has suffered over the

past few years;

123

it does not make sense for them to fail to improve gender

equity in their organizations. As Part VII discusses diversity can be

improved with the responsible development and use of AI.

A. Financial Benefit

A number of studies have demonstrated a link between gender diversity

and corporate performance.

124

For example, Morgan Stanley’s Sustainable +

Responsible Investment (SRI) and Global Quantitative Research teams have

collected and analyzed data on this issue from around the world, and

created a proprietary gender-diversity framework for ranking more than

1,600 stocks globally.

125

The results indicate that a company’s percentage of

female employees is positively correlated with its return on equity.

126

recent study by Quantopian showed that women-led Fortune 1000

companies posted greater cumulative returns than those of the S&P 500,

with an even more pronounced difference after the financial crisis of

2008.

127

The equity returns of the female-led companies were 226% higher

than the S&P 500’s.

128

The opposite holds true at the other end of the

spectrum. Companies in the bottom 25% in terms of gender and ethnic

diversity were the least likely to record profits higher than the national

industrial average.

129

122

. See Sameen, 25 Cutting Edge Technology Examples You Won’t Believe Exist, LIST

25 (Sept. 13, 2018), https://perma.cc/XDJ2-4MCN (describing innovative technologies).

123

. See Jon Swartz, Regaining Trust Is the No. 1 Issue for Tech in 2019, BARRON’S

(Dec. 28, 2019), https://perma.cc/2GR6-VGPU (stating that Big Tech stock prices

dropped significantly in 2018 due to reports of hate speech, fake news, election

interference, perceived anti-conservative bias, privacy violations, business dealings in

China, and a general loss of trust).

124

. MCKINSEY & CO., WOMEN MATTER: TEN YEARS OF INSIGHTS ON GENDER DIVERSITY (2017),

https://perma.cc/XA74-8GJ7.

125

. Why It Pays to Invest in Gender Diversity, supra note 116.

126

. Id. According to the Credit Suisse Research Institute, companies with at least

one woman on their board of directors outperform those without any women by 26%.

Press Release, Credit Suisse, Large-Cap Companies with At Least One Woman on the

Board Have Outperformed Their Peer Group with No Women on the Board by 26% Over

the Last Six Years, According to a Report by Credit Suisse Research Institute (Jul. 31,

2012), https://perma.cc/M5ZN-264U.

127

. Pat Wechsler, Women-Led Companies Perform Three Times Better Than the S&P

500, FORTUNE (Mar. 1, 2013), https://perma.cc/8BU6-X49P.

128

. Id.

129

. Vivian Hunt et al., Why Diversity Matters, MCKINSEY & CO. (Jan. 2015),

https://perma.cc/FY9E-4R7C.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 312

It should be especially important to tech companies that gender

diversity also leads to greater innovation.

130

A study conducted at the

University of Arizona of Fortune 500 companies found that “companies that

have women in top management roles experience . . . ‘innovation intensity’

and produce more patents—by an average of 20% more than teams with

male leaders.”

131

According to Vivek Wadhwa, a Distinguished Fellow at

Carnegie Mellon University’s College of Engineering, “We are effectively

leaving out half of our population by excluding women from the innovation

economy. Given where technology is headed, with technologies advancing

exponentially and converging, the skills needed to solve the larger problems

require a broad understanding of different fields and disciplines.”

132

B. Increased Numbers of Women in Tech

More women in the ranks leads to more women in leadership roles.

More women in leadership roles leads to more women in the ranks. In

addition to the financial benefit, women leaders are associated with the

hiring of more women throughout the company.

133

Women leaders also

tend to lessen the pay gap between men and women.

134

Women CEOs pay

high-level women employees more than male CEOs do.

135

Because women

tend to support diversity and social responsibility, they also implement

more favorable HR policies

136

which also attract more women to the

industry.

137

There are two ways to get more women into leadership roles:

hire them into the role or promote from within the company. Creating a

larger pool of female employees from which to promote greatly increases

130

. Why It Pays to Invest in Gender Diversity, supra note 116; Cristina Diaz-Garcia et

al., Gender Diversity Within R&D Teams: Its Impact on Radicalness of Innovation, 15

INNOVATION: MGMT., POL’Y & PRAC. 149 (2013).

131

. Yoni Blumberg, Companies With More Female Executives Make More Money—

Here’s Why, CNBC (Mar. 2, 2018), https://perma.cc/BQ56-5YT8.

132

. Houser, supra note 26.

133

. Sue Duke, The Key to Closing the Gender Gap? Putting More Women in Charge,

WORLD ECON. F. (Nov. 2, 2017), https://perma.cc/L8XL-GDY7.

134

. Geoffrey Tate & Liu Yang, Female Leadership and Gender Equity: Evidence from

Plant Closure, 117 J. FIN. ECON. 771 (2015).

135

. Id.

136

. Alison Cook & Christy Glass, Do Women Advance Equity? The Effect of Gender

Leadership Composition on LGBT-Friendly Policies in American Firms, 69 HUM. REL. 1431,

1435 (2016).

137

. Marta Riggins, 5 Meaningful Things Companies Are Doing to Recruit and Retain

Women, LINKEDIN TALENT BLOG (Mar. 8, 2018), https://perma.cc/88R2-FNW6.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 313

the chances of women moving into leadership positions, which will in turn

increase the number of women and workplace conditions for women

overall. Women leaders are also associated with the hiring of more URMs.

138

C. Benefits to Women in Leadership

According to Shivaram Rajgopal, the Vice Dean of Research at Columbia

Business School, “Women in leadership positions serve as a significant

deterrent against a permissive culture towards sexual harassment,”

he told The Christian Science Monitor. “You rarely hear of such issues at

Yahoo! where Marissa Mayer was the CEO . . . . [Facebook’s Mark]

Zuckerberg has [chief operating officer] Sheryl Sandberg to temper the . . .

culture.”

139

There are numerous studies showing that women do tend to

make better managers for a variety of reasons.

140

According to one survey

of 7,280 corporate leaders by Zenger and Folkman, women demonstrated

higher competencies than men in 12 of the 16 leadership categories

surveyed.

141

The two areas in which women outscored men by the highest

percentage were taking initiative and driving for results, both important to

138

. CENTER FOR EMP. EQUITY, IS SILICON VALLEY TECH DIVERSITY POSSIBLE NOW? 15 (2018),

https://perma.cc/5US2-9FXX. A study by PwC showed that 42% of female board

directors considered racial diversity to be important compared to only 24% of male

directors. Anne Fisher, Would Your Company Be Any Different If It Had More Women on

the Board?, FORTUNE (May 27, 2015), https://perma.cc/LA3Q-6RBD (citing PWC, 2014

ANNUAL CORPORATE DIRECTORS SURVEY—THE GENDER EDITION (2015)).

139

. Houser, supra note 26.

140

. A meta-analysis of 45 studies comparing the leadership skills of men and

women concluded that women tended to be transformational leaders, while men tended

to be transactional or laissez-faire leaders. Transformational leaders “establish

themselves as role models by gaining followers’ trust and confidence. They state future

goals, develop plans to achieve those goals, and innovate, even when their organizations

are generally successful. Such leaders mentor and empower followers, encouraging them

to develop their full potential and thus to contribute more effectively to their

organizations. By contrast, transactional leaders establish give-and-take relationships

that appeal to subordinates’ self-interest. Such leaders manage in the conventional

manner of clarifying subordinates’ responsibilities, rewarding them for meeting

objectives, and correcting them for failing to meet objectives.” The conclusion of the

meta-analysis was that women are generally more effective leaders, while men are only

somewhat effective or hinder effectiveness. Alice Eagly & Linda L. Carli, Women and the

Labyrinth of Leadership, HARV. BUS. REV., Sep. 2007, https://perma.cc/6ARB-YGQW

(referring to Alice H. Eagly et al., Transformational, Transactional, and Laissez-Faire

Leadership Styles: A Meta-Analysis Comparing Women and Men, PSYCHOL. BULL. (2003)).

141

. Jack Zenger & Joseph Folkman, Are Women Better Leaders than Men?, HARV. BUS.

REV. (Mar. 15, 2012), https://perma.cc/XJF2-STQC.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 314

the tech industry.

142

By promoting women to prominent positions of

leadership, companies may be able to prevent some of the more outrageous

conduct which leads to sexual harassment claims.

143

D. The Need to Fill Tech Jobs in 2020

Perhaps the most important financial reason is to maintain U.S. tech

industry’s dominant position in the world. Currently, there are 11.8 million

people employed in the tech field in the U.S.

144

It is estimated that by 2020

the U.S. will not be able to fill the additional 1 million open positions in

tech.

145

The tech industry needs to start focusing on solutions that work

rather than creating D&I webpages and paying billions to consultants for

training programs that do not work.

146

If tech companies keep limiting their

hiring pool, it will not be possible to fill all of these needed positions. The

top five tech companies in the world are located in the U.S. and account for

45% of the S&P 500’s year-to-date gain.

147

Being unable to hire enough

employees to perform the work required could be catastrophic not just to

these organizations, but to the U.S. economy itself. While the U.S. economy

has grown at a rate of 1.5% per year from 2006-2016, the average annual

growth rate for the digital economy was 5.6% over the same time period.

148

Because the tech industry has been unable to increase the numbers of

women and URMs in any significant way since they began releasing

diversity reports, immediate action needs to take place. The tech industry

seems mystified by the lack of success, but it is not too difficult to figure it

out. What they are doing simply does not work.

142

. Id.

143

. See Frank Dobbin & Alexandra Kalev, Training Programs and Reporting Systems

Won’t End Sexual Harassment. Promoting More Women Will, HARV. BUS. REV. (Nov. 15,

2017), https://perma.cc/RA43-DF9V (stating that “[h]arassment flourishes in

workplaces where men dominate in management and women have little power”).

144

. THE COMPUTING TECH. INDUS. ASSOC., CYBERSTATES 2019 6 (2019),

https://perma.cc/YN4K-QASB.

145

. Alison DeNisco Rayome, CIO Jury: 83% of CIOs Struggle to Find Tech Talent,

TECHREPUBLIC (June 16, 2017), https://perma.cc/SA9M-JNEW.

146

. See Part IV infra.

147

. Jamillah Williams, Diversity as a Trade Secret, 107 GEO. L.J. 1, 5 (2018).

148

. U.S. DEP’T OF COMMERCE, BUREAU OF ECON. ANALYSIS, MEASURING THE DIGITAL ECONOMY:

AN UPDATE INCORPORATING DATA FROM THE 2018 COMPREHENSIVE UPDATE OF THE INDUSTRY

ECONOMIC ACCOUNTS (2019).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 315

V.CURRENT DIVERSITY AND INCLUSION METHODS DO NOT WORK

After Google started issuing its diversity report in 2014, other tech

companies followed suit. In order to address the problem, these companies

spent billions of dollars on consultants and training programs but failed to

increase the numbers of women and minorities as demonstrated in more

recent reports.

149

The most common fix was to institute unconscious bias

workshops and mentoring programs.

A. Training Does Not Work

The primary method put forth for tackling the lack of diversity by tech

companies has been diversity training. The concept behind this idea is that

by explaining to employees and managers their biases, they can actively

avoid them. Although billions have been spent on diversity training, studies

show that it has had no effect in decreasing bias or in increasing diversity in

the companies in which training occurred.

150

Although those taking the

training may be able to complete a quiz, they often quickly forget what they

learned.

151

Unfortunately, not only does diversity training not work,

152

actually can cause more harm.

153

Multiple studies have shown that it has the

149

. According to a Statista report, in 2017 women held 23% of technical positions

at Apple, 17% at Facebook, and 19% at Google. Clare Byrne, 4 Charts That Show Tech is

Still a Man’s World, WORLD ECON. F. (Aug. 7, 2017), https://perma.cc/GPX2-USBZ. This

shows only a slight increase over the 2014 figures of 20%, 15% and 17% respectively.

Felix Richter, Women Vastly Underrepresented in Silicon Valley Tech Jobs, STATISTA

(Aug. 14, 2014), https://perma.cc/CA9B-JQRX. According to data from the Government

Accountability Office, between 2007 and 2015, women in tech positions stood still at

22% on average. Joe Davidson, Mostly White Male Tech Sector Needs Government Help on

Diversity, WASH. POST (Dec. 4, 2017), https://perma.cc/6JEL-P5C8.

150

. Frank Dobbin & Alexandra Kalev, Why Diversity Programs Fail, HARV. BUS. REV.,

July-Aug. 2016, at 52, https://perma.cc/FAA3-6HQH [hereinafter Dobbin & Kalev, Why

Diversity]; Cheryl R. Kaiser et al., Presumed Fair: Ironic Effects of Organizational Diversity

Structures, 104 J. PERSONALITY & SOC. PSYCHOL. 504 (2013); Joanne Lipman, How Diversity

Training Infuriates Men and Fails Women, TIME (Jan. 25, 2018), https://perma.cc/XD5D-

XNS2; Mike Noon, Pointless Diversity Training: Unconscious Bias, New Racism and Agency,

32 WORK, EMP’T AND SOC’Y 198, 203 (2017).

151

. Dobbin & Kalev, Why Diversity, supra note 150, at 54.

152

. An EEOC Task Force found in 2016 that thirty years of social science research

has failed to show that training reduces sexual harassment. Christina Folz, No Evidence

that Training Prevents Harassment, Finds EEOC Task Force, SOC’Y FOR HUM. RES. MGMT.

(June 19, 2016), https://perma.cc/US2L-LKSE; Jena McGregor, To Improve Diversity,

Don’t Make People Go to Diversity Training. Really, WASH. POST (July 1, 2016),

https://perma.cc/7YF8-PXC4.

153

. Susan Bisom-Rapp, Sex Harassment Training Must Change: The Case for Legal

Incentives for Transformative Education and Prevention, 71 STAN. L. REV. ONLINE 62 (2018);

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 316

potential to actually increase bias.

154

In fact, men tend to interpret required

diversity training as an assignment of blame.

155

Instead of encouraging the

equitable treatment of women and minorities, men interpreted the message

as requiring them to provide special treatment to women or minorities or

demonstrated a fear of losing their jobs to women and minorities.

156

Since

2014, some 70,000 of its employees participated in Google’s diversity

training program.

157

However, their 2018 report indicates that the

composition of women barely budged over this timeframe.

158

In fact, this

diversity training led one man to sue for discrimination.

159

Training can cause unanticipated harm to the workplace culture.

Studies shows that training can actually cause women and URMs to believe

that their co-workers are more biased than they actually are.

160

In addition,

research reveals that when employees are told that biases are

“unconscious,” they feel as though they cannot do anything to change their

behavior as bias is just “human nature.”

161

Managers who were told that

stereotypes against women are common, felt more comfortable indicating

Dobbin & Kalev, supra note 150, at 54; Kaiser, supra note 150, at 504. In one study, white

men who participated in a hiring simulation where participants received either a pro-

diversity or neutral message experienced “cardiovascular reactivity” (a negative

physiological response) in response to the diversity message. Tessa L. Dover et al.,

Members of High-Status Groups Are Threatened by Pro-Diversity Organizational Messages,

62 J. EXPERIMENTAL SOC. PSYCHOL. 58 (2016).

154

. Dobbin & Kalev, Why Diversity, supra note 150; Kaiser, supra note 150; Lipman,

supra note 150; Noon, supra note 150.

155

. Lipman, supra note 150.

156

. Id.

157

. David Pierson & Tracey Lien, Diversity Training Was Supposed to Reduce Bias at

Google. In Case of Fired Engineer, It Backfired, L.A. TIMES (Aug. 9, 2017),

https://perma.cc/K4A4-MYNE.

158

. The Google Diversity Report 2014-2018 shows that in 2014 30.6% of its

workforce was female. In 2018, that figure is 30.9% demonstrating no significant

improvement. GOOGLE, GOOGLE WORKFORCE COMPOSITION 2014-2018,

https://perma.cc/P7JH-2Y4S.

159

. “I went to a diversity program at Google and . . . I heard things that I definitely

disagreed with,” said Damore, a former Google employee. “[T]here was a lot of, just,

shaming—‘No, you can’t say that, that’s sexist’; ‘You can’t do this.’” Pierson & Lien, supra

note 157.

160

. Lipman, supra note 150.

161

. Michelle M. Duguid & Melissa C. Thomas-Hunt, Condoning Stereotyping? How

Awareness of Stereotyping Prevalence Impacts Expression of Stereotypes, 100 J. APPLIED

PSYCHOL. 343 (2015); see also Jeffrey Halter, Unconscious Bias: The Fatal Flaw in Diversity

Training, NETWORK EXEC. WOMEN (Jan. 9, 2018), https://perma.cc/YF2L-SN7S (describing

how men leave unconscious bias training thinking that it is a “Get out of Jail Free” card

because they cannot do anything about it).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 317

that they did not want to hire women because the women were “unlikable”

thus giving license to flaunt their biases.

162

One study showed that five years

after diversity training there were 6% fewer black women in management

positions.

163

In addition, the mere existence of a diversity training program

can result in men being less likely to believe that gender discrimination

exists at their company despite evidence to the contrary.

164

B. Mentoring Programs Do Not Work

Tech companies have also tried mentoring programs. Mentoring is

simply a way to shift the burden to women to fix the discrimination they

encounter in the workplace. Teaching women how to negotiate or “lean in”

does not work. In fact, women who try to negotiate salary are viewed as

difficult to work with.

165

Women who are taught how to be leaders are

judged more harshly than their male counterparts exhibiting the same

behaviors.

166

Mentoring is essentially asking those who are marginalized to advocate

for themselves after receiving advice. When that advice is given by a man, it

may not be entirely applicable as men do not face the same issues or have

the same experiences as women do in the workplace. Women mentors may

provide valuable advice, but often, have little influence in making promotion

decisions.

167

In addition, there are simply not enough women in managerial

or leadership positions in these companies to take on the mentoring of early

career women and URMs and doing so puts an additional burden on those

female mentors potentially harming their own careers. Some studies have

shown that women who advocate for diversity are actually penalized in

162

. Lipman, supra note 150; see also Kaiser, supra note 150 at 505 (“[A]sking

people to suppress their stereotypes can inadvertently increase stereotyping and

prejudice”).

163

. Dobbin & Kalev, Why Diversity, supra note 150, at 59.

164

. Kaiser, supra note 150, at 506; Rachel Thomas, How Diversity Branding Hurts

Diversity, MEDIUM (Dec. 7, 2015), https://perma.cc/BFN6-QQQ2.

165

. Bowles, supra note 47.

166

. Snyder, Abrasiveness, supra note 45.

167

. Nancy Novak, Gender Diversity Doesn’t Happen Without Advocacy, CIO (Apr. 16,

2018), https://perma.cc/E6ZA-5L5S.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 318

their performance evaluations.

168

While mentoring can help women feel

included, it has not been shown to advance a woman’s career.

169

These measures have failed because they do not address the underlying

reason for the lack of diversity—subjective decision-making by humans,

which is the rule in 80% of companies.

170

Humans simply cannot make

objective decisions, and this failure harms women and URMs more

significantly than others with respect to employment decisions. Addressing

the issue of bias and noise not only helps to increase diversity, it has been

proven to result in more qualified applicants, better hires, better

promotions, and better retention rates.

171

The following Part explains the

reasons underlying poor decision-making by humans.

VI.UNCONSCIOUS BIAS AND NOISE

Social scientists have discovered that unconscious errors of reasoning

distort human judgments (unconscious bias) and random chance variability

in decisions (noise) occurs more often than people realize.

172

Unconscious

biases occurs when people are unaware of the mental shortcuts they use to

process information.

173

Noise refers to variability in human decision-

making due to chance or irrelevant factors.

174

Daniel Kahneman gives the

example of a faulty scale. If your scale consistently reads 10 pounds heavier

than you know yourself to be, it is biased. If every time you step on the scale

168

. Stefanie K. Johnson & David R. Hekman, Women and Minorities Are Penalized for

Promoting Diversity, HARV. BUS. REV. (Mar. 23, 2016), https://perma.cc/HBK2-M47L.

169

. SYLVIA ANN HEWLITT, FORGET A MENTOR, FIND A SPONSOR THE NEW WAY TO FAST-TRACK

YOUR CAREER (2013).

170

. Supra note 8, at 11.

171

. See Part VI infra.

172

. See Jim Holt, Two Brains Running, N.Y. TIMES (Nov. 25, 2011),

https://perma.cc/UAS7-JK2L (discussing Kahneman and Tversky’s series of

experiments that “revealed twenty or so ‘cognitive biases’—unconscious errors of

reasoning that distort our judgment of the world.”); J. Nathan Matias, Bias and Noise:

Daniel Kahneman on Errors in Decision-Making, MEDIUM (Oct. 17, 2017),

https://perma.cc/BSX7-YYF3 (discussing Kahneman’s series of experiments at an

insurance company that revealed unexpected variance in decisions of different

underwriters as well as within the individual underwriter him- or herself).

173

. PARADIGM, MANAGING UNCONSCIOUS BIAS—STRATEGIES TO ADDRESS BIAS & BUILD MORE

DIVERSE, INCLUSIVE ORGANIZATIONS 2 (2016), https://perma.cc/6STW-UUF9.

174

. Daniel Kahneman et al., Noise: How to Overcome the High, Hidden Cost of

Inconsistent Decision Making, HARV. BUS. REV., Oct. 2016, https://perma.cc/CLF8-WJEM

[hereinafter Kahneman et al., Noise: How to Overcome].

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 319

you get a different number, it is noisy.

175

Although people believe they are

objective, noise and bias result in inaccurate and inconsistent decisions.

176

Prior to discussing how the responsible development and use of AI can

mitigate these human errors impacting employment decisions, the

behavioral science background on the many flaws in human judgment will

be explored.

A. Unconscious Bias

In Thinking, Fast and Slow, Kahneman explains how our brains operate

using two systems: System 1 (Fast Thinking) operates automatically (such

as driving a car), whereas System 2 (Slow Thinking) involves reasoning

(such as solving a math problem).

177

Both systems have deficiencies

impacting our ability to make objective decisions. The main problem with

System 1 thinking is it is not prone to doubt or careful reflection because it

involves mental short-cuts to reach a conclusion. This means we believe the

answers we come up with using System 1 thinking are accurate and we do

not spend time analyzing how we got there, resulting in unjustified

confidence in our System 1 decisions. This in turn leads to judgments which

are neither accurate nor logical but become endorsed by System 2 and turn

into deep-rooted beliefs.

178

While Kahneman goes on to explain the

potential errors in judgment when people rely on System 1 thinking, the

most important takeaway is that not only are our judgments untrustworthy,

but the more decisions we make based on these biases, the stronger the

neural connections become that reinforce our belief in the biases’

conclusion. These cognitive biases affect opinions on social issues (as we can

see with prejudice) and hence affect social institutions (such as the

workplace) that rely on individuals to make rational (unbiased) judgments.

Social science literature regarding unconscious bias has established

that we are generally unaware of our own prejudices. We routinely make

decisions and judgments based on incomplete information/knowledge and

a lack of objective truth. We use “mental shortcuts,” known as heuristics, to

fill in gaps in our knowledge with similar data from past experiences and

175

. Id.

176

. MAHZARIN BANAJI & ANTHONY G. GREENWALD, BLINDSPOTS: HIDDEN BIASES 152 (2016).

177

. DANIEL KAHNEMAN, THINKING, FAST AND SLOW 20-22 (2011) [hereinafter KAHNEMAN,

THINKING].

178

. Ameet Ranadive, What I Learned from “Thinking, Fast and Slow,” MEDIUM

(Feb. 20, 2017), https://perma.cc/32KW-X7HJ.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 320

cultural norms. The notion of cognitive biases was introduced to the modern

era by Amos Tversky and Daniel Kahneman in 1972.

179

Essentially, Tversky

and Kahneman exposed that while heuristics are useful, they also lead to

errors in thinking.

180

Their theories also explain why the tech industry

seems unable to correct its diversity problem. Representative bias, for

example, occurs when people conclude that someone works in a given

profession due to how similar that person appears to a stereotype.

181

When

a hiring agent internally views their ideal applicant as a young white male

from an Ivy League school, this stereotype limits their ability to value other

types of candidates.

The reason employment interviews are still handled by humans with

biases is due to the validity illusion. As Kahneman and Tversky explain,

people tend to overrate their own ability to make accurate predictions.

182

This validity illusion exists because of confirmation bias—focusing on

information that fits our prediction and discarding that which does not.

183

Therefore, during an interview, the interviewer, without realizing it, may

notice only the information that confirms their pre-existing belief rather

than the information relevant to the position.

184

179

. Daniel Kahneman & Shane Frederick, Representativeness Revisited: Attribute

Substitution in Intuitive Judgment, in HEURISTICS AND BIASES: THE PSYCHOLOGY OF INTUITIVE

JUDGMENT 49, 51-52 (Thomas Gilovich, Dale Griffin, & Daniel Kahneman eds., 2002).

180

. Amos Tversky & Daniel Kahneman, Judgment Under Uncertainty: Heuristics and

Biases, 185 SCI. 1124, 1124 (1974) [hereinafter Tversky & Kahneman, Judgment].

181

. Id. at 1124. For example, if someone’s experience as a child was to see male

doctors and female nurses during office visits, when they hear the word “doctor” an

image of a male will immediately come to mind. Recall the riddle: “[A] father and son are

in a horrible car crash that kills the dad. The son is rushed to the hospital; just as he’s

about to go under the knife, the surgeon says, ‘I can’t operate—that boy is my son!’

Explain.” A study at BU found that most of the participants could not answer because

they did not consider the possibility that the surgeon was the boy’s mother. Rich Barlow,

BU Research: A Riddle Reveals Depth of Gender Bias, BU TODAY (Jan. 16, 2014),

https://perma.cc/8WL2-WL2S.

182

. Daniel Kahneman & Amos Tversky, On the Psychology of Prediction, 80 PSYCHOL.

REV. 237, 249 (1973).

183

. Raymond S. Nickerson, Confirmation Bias: A Ubiquitous Phenomenon in Many

Guises, 22 REV. GEN. PSYCHOL. 175, 175-76 (1998) (explaining scholarly definitions of

confirmation bias, each line of study emphasizing different features of behavior).

184

. Laszlo Block, Here’s Google’s Secret to Hiring the Best People, WIRED (Apr. 7,

2015), https://perma.cc/Q5Z8-E8F4 (stating that a University of Toledo study

concluded “most interviews are a waste of time because 99.4 percent of the time is spent

trying to confirm whatever impression the interviewer formed in the first ten seconds”).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 321

We tend to prefer people who look like us, think like us and come from

backgrounds similar to ours.

185

This is known as the affinity bias

186

or “in-

group favoritism.”

187

One Silicon Valley investor explained that Silicon

Valley’s diversity problem comes from mistaking certain traits, such as

personality type or alma mater, for actual skills: “You don’t necessarily have

to be biased against somebody. You can be biased in favor of somebody.”

188

This favoritism is “responsible for many discriminatory outcomes, from

relying on homogenous employee referral networks for new hires to giving

more favorable performance reviews to employees with similar personality

traits as you.”

189

In fact, part of the problem is that not only are these biases

unknown to those who possess them, but they are particularly difficult for

individuals and organizations to correct.

190

While we believe ourselves to be open-minded and objective, research

shows that the beliefs and values acquired from family, culture and a

lifetime of experiences heavily influence how we view and evaluate others.

Stereotyping involves making assumptions based on the group an

individual belongs to, which has enormous implications for women working

in male dominated fields. One study reported how potential employers

systematically underestimated the mathematical performance of women,

which resulted in hiring less qualified males over higher-performing

women.

191

A recent study of college students also discloses that males tend

to believe they are smarter than women (they were about two times as

185

. See Caren B. Goldberg, Relational Demography and Similarity-Attraction in

Interview Assessments and Subsequent Offer Decisions: Are We Missing Something?, 30 GRP.

& ORG. MGMT. 597, 598 (2005) (explaining, and then complicating, the prevailing view

that recruiters favor applicants similar to themselves).

186

. ELIZABETH LINOS & JOANNE REINHARD, CHARTERED INST. OF PERSONNEL & DEV., A HEAD

FOR HIRING: THE BEHAVIOURAL SCIENCE OF RECRUITMENT AND SELECTION, at 7, 24 (2015).

187

. Marilynn B. Brewer, The Importance of Being We: Human Nature and Intergroup

Relations, 62 AM. PSYCHOL. 728, 729 (2007).

188

. Tracy Ross, The Unsettling Truth About the Tech Sector’s Meritocracy Myth,

WASH. POST (Apr. 13, 2016), https://perma.cc/8EDF-6WDX.

189

. Id.

190

. Valerie Martinelli, The Truth About Unconscious Bias in the Workplace,

TALENTCULTURE (Mar. 31, 2017), https://perma.cc/J7KS-NZ5B.

191

. CHRISTIANNE CORBETT & CATHERINE HILL, AM. ASS’N OF U. WOMEN, SOLVING THE

EQUATION: THE VARIABLES FOR WOMEN’S SUCCESS IN ENGINEERING AND COMPUTING 3 (2015). Even

the President of Harvard University publicly attributed the lack of females in science and

math at top universities to innate differences between the sexes; although Larry

Summers apologized, he resigned shortly thereafter. Alan Finder et al., President of

Harvard Resigns, Ending Stormy 5-Year Tenure, N.Y. TIMES (Feb. 22, 2006),

https://perma.cc/AJ48-38ND.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 322

likely to report this).

192

This is especially ironic given that for a number of

years colleges have been excluding well-qualified female applicants from

college admissions, preferring instead the less-qualified male applicants in

order to achieve gender parity.

193

Status quo bias exists when people prefer situations to remain the

same.

194

Thus, if the workplace is primarily male and has been for a long

time, there is an inherent bias against introducing greater numbers of

women into that environment because it will change how “things have

always been done.” Part of the reason for this bias is loss aversion. People

weigh the potential for loss more heavily than the potential for gain. Even

when the workplace will gain advantages to the increase in the number of

women, status quo bias will cause people to bristle against the idea because

they fear change. Another reason is the exposure effect. When you have

been exposed to a workplace that is predominantly male over a long period

of time, it becomes a preference. Malcolm Gladwell also describes research

from psychology and behavioral economics revealing mental processes that

work rapidly and automatically from relatively little information.

195

Unfortunately, snap judgments can be the result of subconscious racial or

gender bias. Gladwell explains that prejudice can operate at an intuitive

unconscious level, even in individuals whose conscious attitudes are not

prejudiced.

196

Another issue is that humans do not make consistent

decisions.

192

. See Katelyn M. Cooper et al., Who Perceives They Are Smarter? Exploring the

Influence of Student Characteristics on Student Academic Self-Concept in Physiology, 42

ADVANCES PHYSIOLOGICAL EDUC. 200, 205 fig.2 (2018).

193

. Marcus B. Weaver-Hightower, Where the Guys Are: Males in Higher Education,

42 CHANGE: MAG. HIGHER LEARNING, July 8, 2010, at 29, 30.

194

. William Samuelson & Richard Zeckhauser, Status Quo Bias in Decision-Making,

1 J. RISK & UNCERTAINTY 7, 8 (1988). Samuelson & Zeckhauser give the example of choosing

a sandwich for lunch that you have once had before because of the perceived risk in

choosing a different sandwich that you might not like. Id. at 10. In addition, they point to

the popular example of “New Coke” being the preferred taste in a blind test but not

preferred in the marketplace where consumers see and prefer the Coke with which they

are familiar. Id. at 11.

195

. See, e.g., MALCOLM GLADWELL, BLINK: THE POWER OF THINKING WITHOUT THINKING 8-11

(2005).

196

. Id. at 96-97.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 323

B. Noise

According to Kahneman, although organizations expect consistency in

decision-making, this is seldom the case. “The problem is that humans are

unreliable decision makers; their judgments are strongly influenced by

irrelevant factors, such as their current mood, the time since their last meal,

and the weather. This variance in decisions is known as noise.”

197

Research

demonstrates that even looking at identical information, managers’

decisions will vary from those of other managers. Noise also occurs when

managers make decisions inconsistent with their prior decisions.

198

This

inconsistent decision-making costs organizations billions in lost

productivity, and exposes them to potential liability.

199

In making employment decisions, not only do we want to eliminate or

reduce unconscious biases, we also want to eliminate or reduce noise.

Consistent decisions are more equitable and can help avoid claims of

discrimination.

200

The difficulty is that not only are human decision-makers

unaware of the bias and noise in their decisions,

201

these problems may not

be detectable by other humans.

202

Because professionals believe they can

accurately predict who will make a good hire for a particular position, they

are not likely to believe that they are inconsistent in their own decision-

making or that their decisions vary significantly from those of their

colleagues.

203

Relying on humans to make employment decisions produces

not only biased and inconsistent results, but also less accurate ones.

Kahneman and others have suggested incorporating AI into the decision-

making process can mitigate the impact of illogical human decisions.

204

197

. Kahneman et al., Noise: How to Overcome, supra note 174.

198

. Id.

199

. Id.

200

. See Part VIII infra.

201

. Mendez, supra note 82.

202

. Kahneman et al., Noise: How to Overcome, supra note 174.

203

. Guide: Use Structured Interviewing, GOOGLE: RE:WORK, https://perma.cc/HDB3-

XLRQ (archived Apr. 10, 2019). This overconfidence bias increases with the addition of

more information. Rather than increasing accuracy, human decision-makers fail to

adjust for their cognitive limitations. Claire I. Tsai et al., Effects of Amount of Information

on Judgment Accuracy and Confidence, 107 ORG. BEHAV. & HUM. DECISION PROCESSES 97, 97

(2008).

204

. James Pethokoukis, Nobel Laureate Daniel Kahneman on AI: ‘It’s Very Difficult to

Imagine That with Sufficient Data There Will Remain Things That Only Humans Can Do’,

AEIDEAS (Jan. 11, 2018), https://perma.cc/NP4P-5GMY.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 324

VII.USING AI TO REDUCE BIAS/NOISE IN HUMAN DECISION-MAKING

The reasons for the lack of women in the tech industry can be attributed

to several factors, lower levels of hiring and promotion, unfriendly

workplace-culture, and the inability of tech companies to keep the women

they do hire. Ultimately, the underlying cause of the lack of diversity stems

from flawed human decision-making, especially by managers and others

who influence the hiring, promotion, and retention of women. As explained

in the previous Part, human decisions are fraught with unjustified

variability and bias. It is possible, however, to circumvent many of the

problems inherent in faulty human decision-making through the

responsible use of AI.

205

As mentioned in the Introduction, an algorithm is a

series of rules programmed into a computer, while AI is a broader term

covering the process of using a machine to perform a function formerly

performed by humans, such as conducting word searches, testing for skills,

and analyzing data and producing outcomes.

206

Some utilizations of AI

include anonymizing resumes and interviewees, performing structured

interviews through online submission or chatbots, parsing job descriptions,

using neuroscience games which can identify traits, skills and behaviors

which can then be used to match candidates with open positions, predicting

which employees are looking to leave to improve retention, mining

employee reviews for biased language, and standardizing promotion

decisions.

A. Tackling Unconscious Bias

Companies that have moved from traditional recruiting methods to

using AI have found success in creating a more diverse slate of

candidates.

207

One method shown to increase the diversity of candidates is

205

. EXEC. OFFICE OF THE PRESIDENT, BIG DATA: A REPORT ON ALGORITHMIC SYSTEMS,

OPPORTUNITY, & CIVIL RIGHTS 14 (2016), https://perma.cc/MEK5-8EBT. The report can be

read as contemplating only rule-based, algorithmic decision-making using big data—so-

called good-old fashion artificial intelligence—but its basic point applies equally to the

opportunity for machine learning using big data.

206

. Donahue, supra note 14. Data mining is the process of a machine reviewing a

data set to find patterns and relationships between variables, such as what type of coding

skills translate to good performance in a particular position. Machine learning is the

ability of a machine to improve its performance of an outcome by modifying the

algorithm on its own without the intervention of the programmer.

207

. Pymetrics, for example, reports that traditional resume-reviewing results in

women and URMs being at a 50-67% disadvantage while companies reported an

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 325

the use of algorithms to remove race, gender, and national origin from the

initial evaluation process.

208

For example, Unbias.io, removes faces and

names from LinkedIn profiles to reduce the effects of unconscious bias in

recruiting, while Interviewing.io eliminates unconscious bias by providing

an anonymous interviewing platform.

209

Another company, Entelo,

anonymizes interviewing by removing all indication of gender or race.

210

Talent Sonar writes gender-neutral job descriptions and hides applicants’

names, gender, and other personal identifiers from those doing the

hiring.

211

Textio, a program that rewords job ads to appeal to a broader

demographic, managed to successfully increase the Australian software

company Atlassian’s percentage of women among new recruits from 18%

to 57%.

212

Removing names and gender identifications from resumes results in the

hiring of more women.

213

One problem is that although organizations will

increase in diversity of 20-100% of their hires. PYMETRICS: EMPLOYERS,

https://perma.cc/JL3X-NYWE (archived May 21, 2019).

208

Interviewing.io allows candidates to practice interviewing with former

executives at tech companies. When a candidate becomes proficient at the

practice interviews, they can be invited to interview anonymously at tech

companies and can skip directly to the tech phase of the interview—that is, the

part of the interview where the candidate is tested on solving algorithmic

coding problems or some other technical problem. In other words,

Interviewing.io allows people to skip the initial in-person screening that is

currently one point at which bias can creep in during traditional interviews.

Finally, if a candidate feels they have done well at the tech interview, they can

choose to “unmask” themselves—at which point the next phase is generally an

onsite interview. So, while the face-to-face interaction that sometimes triggers

bias does still occur with Interviewing.io, it occurs at a much later stage. By this

time many candidates will have already demonstrated facility for many of the

tasks associated with the job and perhaps any bias is reduced or counteracted

by awareness of their technical skill.

Nancy Leong, The Race Neutral Workplace of the Future, 51 U.C. DAVIS L. REV. 719, 724

(2017).

209

. Id.

210

. “The software allows recruiters to hide names, photos, school, employment

gaps and markers of someone’s age, as well as to replace gender-specific pronouns—all

in the service of reducing various forms of discrimination.” Rebecca Greenfield & Riley

Griffin, Artificial Intelligence Is Coming for Hiring, and It Might Not Be That Bad, L.A. TIMES

(Aug. 10, 2018), https://perma.cc/RK3J-BS8V.

211

. Alsever, supra note 84.

212

. Simon Chandler, The AI Chatbox Will Hire You Now, WIRED (Sept. 13, 2017),

https://perma.cc/QN6T-WZFZ.

213

. Claire Cain Miller, Is Blind Hiring the Best Hiring?, N.Y. TIMES MAG. (Feb. 28,

2016), https://perma.cc/T76G-ZXQX. One study presented at the National Academy of

Sciences in 2012 found that “female student[s] [were judged] to be less competent and

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 326

hire and promote men based on their potential, women tend to be hired and

promoted only on proven past performance.

214

Even when resumes are

identical, organizations choose a man for an interview more often than the

women with identical credentials.

215

By hiding the attributes of the

applicants that could give rise to biased assumptions, the possibility of

discrimination on this basis is removed.

216

Formalized hiring and promotion criteria also help reduce

subjectivity.

217

Slack, for example, uses “white board interviews” where

candidates solve problems at home. Organizations can prevent bias by first

removing a candidates’ identifying information, and then evaluating the

candidates’ work against a comprehensive checklist.

218

Structured

interviews are another proven way to eliminate or reduce bias.

219

During a

less worthy of being hired than identical male student[s], and also [were] offered a

smaller starting salary and less career mentoring.” Corinne A. Moss-Racusin et al., Science

Faculty’s Subtle Gender Biases Favor Male Students, 109 PROC. NAT’L ACAD. SCI., 16474,

16477 (2012).

214

. Carter & Silva, supra note 58.

215

. Rhea E. Steinpreis, Katie A. Anders, & Dawn Ritzke, The Impact of Gender on the

Review of the Curricula Vitae of Job Applicants and Tenure Candidates: A National

Empirical Study, 41 SEX ROLES 509, 522-24 (Oct. 1999).

216

. Alsever, supra note 84.

217

. Eric Luis Uhlmann & Geoffrey L. Cohen, Constructed Criteria Redefining Merit to

Justify Discrimination, 16 PSYCHOL. SCI. 474, 474 (2005). One study at Yale showed that

employers often justify bias after the fact. When those evaluating candidates for a police

chief position saw a female name for the candidate who had a degree and a male name

for the candidate who did not, the evaluators indicated that they chose the man because

“street smarts” were the most important factor. When the names were reversed, the

evaluators justified choosing the man because the degree was the most important factor.

However, when the hiring criteria was formalized before they looked at applications (i.e.

reviewers decided in advance whether formal education or street smarts was more

important), bias was reduced and the candidate matching the criteria was more likely to

be chosen. Id.

218

. Lauren Romansky & Emily Strother, Slack’s Unique Diversity Strategy Offers

Some Lessons for Silicon Valley and Beyond, TALENT DAILY (May 15, 2018),

https://perma.cc/J452-JCTG. Interview.io uses software that disguises voices so that the

interviewer cannot determine the sex of the interviewee. Slack has successfully

increased its diversity through AI measures such as the ones described in this paper.

Slack employs 44.7% women and an independent third party confirmed Slack’s pay

equity achievement. Their workforce consists of 12.6% underrepresented minorities

and 8.3% identify as LGBTQ. Diversity at Slack: An Update on Our Data, April 2018, SLACK

BLOG (Apr. 2018), https://perma.cc/XK5T-UBDG. Another future fix suggested by Leong

is to incorporate virtual reality offices where employees interact with one another via a

virtual world where employees could choose any avatar and everyone would understand

that it would not necessarily reflect their gender, race or national origin. Leong, supra

note 208.

219

. Michael McDaniel et al., The Validity of Employment Interviews: A Comprehensive

Review and Meta-Analysis, 79 J. APPLIED PSYCHOL. 599, 599 (1994).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 327

structured interview, each candidate answers questions identical to those

asked of the other interviewees. According to Loren Larsen, CTO of HireVue,

“[b]y using a structured interview process where the questions are carefully

designed and the answers are analyzed and compared with hundreds or

thousands of samples, we can begin to predict job performance more

accurately than human evaluators and begin to remove bias that has always

been present in the hiring process.”

220

Mya Systems created a chatbot that

recruits, interviews, and evaluates job candidates using performance-based

questions.

221

The chatbot compares the answers with the job requirements

and answers the candidates’ questions about the position and company.

222

This technological capability permits the initial evaluation of the candidate

to be based on predetermined criteria without human biases creeping in.

Fortune magazine has identified some 75 startups entering the field of

AI hiring programs.

223

Not only can companies use AI to anonymize

candidates; it can also discover desired attributes in candidates and

employees.

224

Although Amazon was not successful in this regard, as

discussed in Part VII infra, Pymetrics has succeeded in increasing gender

diversity through AI. In addition to creating custom unbiased gamified

assessments of candidates, they have also prevented bias from creeping in

by continually auditing their own algorithms for biased outcomes.

225

The

way Pymetrics works is that candidates engage in neuroscience games

which allows the algorithm to match candidates with positions based on

identified factors.

226

It creates an algorithm to match candidates with its

client’s open jobs. Unlike traditional pre-employment tests, in which

candidates can create answers based on what they think the employer is

looking for, games test actual behaviors, skills and traits shown to be

220

. Melissa Locker, How to Convince a Robot to Hire You, VICE (Oct. 17, 2018),

https://perma.cc/LEF8-3DV5.

221

. Chandler, supra note 212.

222

. Id.

223

. Alsever, supra note 84.

224

. Robert Bolton, Artificial Intelligence: Could an Algorithm Rid Us of Unconscious

Bias?, PERSONNEL TODAY (Nov. 16, 2017), https://perma.cc/4RSN-FTB4 (explaining that

an algorithm can be designed to assess “performance against a predetermined

personality profile” rather than asking someone if they have the right experience).

225

. Greenfield & Griffin, supra note 210.

226

. PYMETRICS, supra note 207. Pymetrics’ technology can also match candidates

with other positions should they not have the skills needed for the job they are interested

in.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 328

associated with top performers in the company.

227

Unilever reported that

since they began using Pymetrics, they have doubled the number of

applicants they hire after the final round of interviews, increased revenue

by hiring a better quality of employee, and increased the diversity of their

applicant pool.

228

One of the advantages of moving to online assessments and games is the

ability to locate non-traditional applicants. Many people with technical

skills have not gone to college or have left the workforce for a period of time.

Traditional resume screening can eliminate qualified candidates by

discarding those without a college degree or who have a large gap in their

work history. By focusing on the skills rather than pedigree of applicants,

companies can not only locate more qualified employees best suited to the

position, but also a more diverse group. HackerRank notes that because

65% of programmers are at least partially self-taught,

229

if a company is

conducting their own search on college campuses, they are likely choosing

from a smaller, wealthier, homogenous set of candidates and ignoring a

large pool of qualified, more diverse candidates. HackerRank allows anyone

to go online and participate in coding and other technical assessments so

that the hiring company can assess the applicant’s skills, rather than

pedigree.

230

GapJumpers reports that their skills-testing AI resulted in 60%

of the women and URMs applying getting an initial interview, up from 20%

with resume screening.

231

Another subcomponent of AI, data mining, involves the discovery of

patterns and relationships in a data set.

232

Data mining can be used to

227

. Chris Ip, To Find a Job, Play These Games, ENGADGET (May 4, 2018),

https://perma.cc/Z3TS-MQYD.

228

. Wanda Thibadeaux, Unilever Is Ditching Resumes in Favor of Algorithm-Based

Sorting, INC. (June 28, 2017), https://perma.cc/3XVC-4XY2. Infor reports boosting

employee diversity for some of its clients by 26%, and Famous Footwear achieved a 33%

lower turnover rate after adopting AI in its hiring process. Jill Strange, Cut Out the

Unconscious Bias in Your Hiring with Smart Data, DIGINOMICA (Jul. 10, 2017),

https://perma.cc/3BJE-BXXL. Another company, Stella, also matches applicant with

potential positions through the use of AI. Like Pymetrics, it also audits their algorithms

to detect bias. Greenfield & Griffin, supra note 210.

229

. VIVEK RAVISANKAR, HACKERRANK, STUDENT DEVELOPER REPORT 2018 2 (2018),

https://perma.cc/BVQ4-R84S.

230

. I refer to specific companies in this paper as examples of the wide range of

services available today. For a more inclusive list, see Kayla Kozan, The 38 Top Recruiting

Software Tools of 2019, IDEAL BLOG (Feb. 4, 2019), https://perma.cc/DK4X-3ML7.

231

. Miller, supra note 213.

232

. Data Mining: What It Is and Why It Matters, SAS INSIGHTS,

https://perma.cc/5C85-BSKF (archived May 15, 2019).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 329

extract information from large data sets to identify which factors are

associated with retention.

233

Data mining can discover correlations,

patterns, and variances within huge data sets to predict outcomes.

234

The

following examples used internal employee data for the analysis, not data

mined from the internet.

235

Retention is an important part of keeping the

women that are hired in tech. In an early study performed by Chien and

Chen (2008), they successfully demonstrated how using “decision tree

analysis to discover latent knowledge and extract the rules” to assist in

personnel selection decisions and generate recruiting and human resource

management strategies.

236

AI can analyze employee data to reduce turnover, increase efficiency,

improve employee engagement, and predict job performance. Sysco’s AI

program improved its retention rate from 65% to 85% by tracking

employee satisfaction scores, which allowed them to institute immediate

improvements and saved Sysco nearly $50 million in hiring and training

costs for new associates.

237

AI advances in recent years along with the

availability of lower cost cloud storage has vastly improved the ability of

machines to analyze large data sets.

238

Deloitte’s Bersin Study on People

Analytics indicates that the use of AI strongly correlates with improved

233

. EXEC. OFFICE OF THE PRESIDENT, supra note 205.

234

. SAS INSIGHTS, supra note 232.

235

. Data mined from the internet or social media presents significant problems that

technology, at least today, cannot fully address. Companies such as IBM and Google are

working on software that can flesh out and correct bias in large data sets, but for the

purposes of this paper, data mining examples are limited to known datasets.

236

. Chen-Fu Chien & Li-Fei Chen, Data Mining to Improve Personnel Selection and

Enhance Human Capital: A Case Study in High-Technology Industry, 34 EXPERT SYS. WITH

APPLICATIONS 280, 281 (2008). A similar study was performed in 2013 by Azar et. al with

respect to the banking industry. The study used data mining and decision tree analysis

to create recruiting rules to improve employee retention. After examining 26 variables,

only five were found to impact job performance. This time, gender was identified and

found to have no impact on “upgrade status.” The research design expressly indicated

that “[t]he present thesis takes this approach further by avoiding opinion-based methods

that are traditionally used in the selection of new employees” (emphasis added). The

study concluded that data mining techniques are an extremely important tool to help

managers discover covert knowledge, which can assist in human resource management.

Adel Azar et al., A Model for Personnel Selection with a Data Mining Approach: A Case Study

in a Commercial Bank, 11 SA J. HUM. RES. MGMT. 449, 449 (2013).

237

. Thomas H. Davenport, Jeanne Harris, & Jeremy Shapiro, Competing on Talent

Analytics, HARV. BUS. REV., Oct. 2010, at 54, 56, https://perma.cc/BS8T-86KE.

238

. AI is being adopted more frequently because of the increase in cloud-based HR

systems. Dimple Agarwal et al., People Data: How Far Is Too Far?, in THE RISE OF THE SOCIAL

ENTERPRISE: 2018 DELOITTE GLOBAL HUMAN CAPITAL TRENDS 89, 89 (2018).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 330

talent outcomes and profitability.

239

The report also noted that 40% of

companies utilize cloud-based HR management systems allowing for easier

analytics. The availability of open source AI permits companies to customize

their algorithms.

240

AI can be used to monitor managers as well. AI can search for bias in

employee reviews by scanning for certain words to describe female

employees that are not used to describe males, such as “aggressive” for

female, where the same traits would be described as “leadership-material”

for a male.

241

Using AI can directly address the problem with women being

judged more harshly than men.

242

In addition to text mining performance

reviews, SAP has developed algorithms to help companies review job

descriptions for biased language.

243

Overall, responsible AI has been very

successful in reducing bias and increasing diversity.

B. Reducing/Eliminating Noise

Kahneman’s most recent area of research examines how variability in

decisions (known as noise) can negatively impact organizations and how AI

can reduce and or eliminate this variability. AI can help managers and

employers make better decisions by mitigating human biases and

decreasing the inevitable variability in decision-making.

244

Not only are

humans inconsistent in their own decision-making from day to day,

inconsistent decisions also result from two different humans looking at the

same data.

245

By contrast, an algorithm will always provide the same

239

. Deloitte’s Bersin Finds Effective Use of People Analytics Is Strongly Related to

Improved Talent and Business Outcomes, PR NEWS WIRE (Nov. 13, 2017),

https://perma.cc/8X75-JC4Z.

240

. Id.

241

. Courtney Seiter, Bias at Work: Three Steps to Minimizing Bias in Performance

Reviews, BUFFER OPEN BLOG (Feb. 8, 2018), https://perma.cc/5VWK-8PMB.

242

. Moss-Racusin, supra note 53, at 16,474-79 (stating that science faculty rated

male applicants significantly more competent than female applicants with identical

application materials).

243

. Tami Reiss, Big Data to Help Overcome Unconscious Bias—Thanks SAP!, MEDIUM

(May 19, 2016), https://perma.cc/B5JG-59Y3.

244

. Kahneman et al., Noise: How to Overcome, supra note 174. NOISE, a new book

coming out in 2020 or 2021 by Daniel Kahneman, Olivier Sibony, and Cass R. Sunstein,

will explain how variability in human decisions occurs and offer solutions. I anticipate

there will be a great influx of research on noise and algorithms coming out after this book

is published.

245

. Id. at 40.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 331

decision for the same data set.

246

Creating rules that are consistently

applied to data sets reduces noise and creates uniformity throughout the

organization.

247

The use of AI will also achieve greater accuracy than the use

of human decision-makers and avoid decisions which treat similarly

situated applicants and employees differently.

248

According to Kahneman,

“An algorithm could really do better than humans, because it filters out

noise. If you present an algorithm the same problem twice, you’ll get the

same output. That’s just not true of people.”

249

Machine decision-making

outcomes are not only more consistent, they also tend to be more accurate

than those made by humans.

250

Kahneman’s study on insurance underwriters revealed that the

organizations involved were unaware of the variability in their risk

assessment determinations, which Kahneman found to vary on average by

48% in company A and 60% in company B.

251

Kahneman recommended the

use of algorithms to reduce such noise in decision-making.

252

According to

Kahneman, “It has long been known that predictions and decisions

generated by simple statistical algorithms are often more accurate than

those made by experts, even when the experts have access to more

246

. Kahneman also provides a method for auditing decisions for noise. Id. at 42.

247

. Kahneman explains that large data sets are not necessary, the creation of rules

based on set criteria is key. Id. at 46.

248

. Kahneman explains that decisions made by algorithms are “often more accurate

than those made by experts, even when the experts have access to more information than

the formulas use.” Id. at 41.

249

. MIT Initiative on Digital Econ., Where Humans Meet Machines: Intuition,

Expertise and Learning, MEDIUM (May 18, 2018), https://perma.cc/66VT-6A84.

250

. This is especially true in medical testing. See Yun Liu et al., Artificial Intelligence-

Based Breast Cancer Nodal Metastasis Detection: Insights Into the Black Box for

Pathologists, 143 ARCHIVES PATHOLOGY & LABORATORY MED. 859, 861-62 (2018) (noting that

Google researchers’ AI to detect metastatic breast cancer by evaluating lymph node

slides was able to detect 99.3% of the cancers, while human pathologists were only able

to detect 81%); Kazimierz O. Wrzeszczynski et al., Comparing Sequencing Assays and

Human-Machine Analyses in Actionable Genomics for Glioblastoma, 3 NEUROLOGY GENETICS,

Aug. 2017, e164 at 1, 6 (discussing IBM’s Watson, known as WGA (Watson Genomic

Analytics), which was able to analyze a genome of a patient in 10 minutes compared to

the human experts who took 160 hours); European Lung Found., AI Improves Doctors'

Ability to Correctly Interpret Tests and Diagnose Lung Disease, SCIENCEDAILY (Sept. 18,

2018), https://perma.cc/TJJ5-2H3J (comparing AI’s 100% compliance rate for

pulmonary function tests (PFTs) with human pulmonologists’ 74% compliance rate, and

noting further that AI provided a correct diagnosis 82% of the time compared with 45%

by the pulmonologists).

251

. Kahneman et al., Noise: How to Overcome, supra note 174, at 42.

252

. Elisabeth Goodman, Decision Making: Noise, Intuition and the Value of Feedback,

ELISABETH GOODMAN’S BLOG (Feb. 1, 2017), https://perma.cc/6A8H-NALU.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 332

information than the formulas use. It is less well known that the key

advantage of algorithms is that they are noise-free: Unlike humans, a

formula will always return the same output for any given input. Superior

consistency allows even simple and imperfect algorithms to achieve greater

accuracy than human professionals.”

253

One example of an algorithmic fix for reducing noise in hiring is the use

of structured interviews. By using chatbots to conduct interviews

companies eliminate the variability that a human interviewer or multiple

interviewers would bring into the process. Unlike humans, a chatbot asks

each interviewee the same set of questions. In addition, by setting criteria

for promotions in advance, using an algorithm to assess employees will

reduce both the bias of managers and noise in the decision by applying rules

uniformly.

254

VIII.USING AI TO REDUCE ALGORITHMIC BIAS

A number of scholars have written about the dangers and risks of

various subcomponents of AI, such as big data, machine learning, data

mining and predictive analytics.

255

In their paper, Big Data’s Disparate

Impact, Barocas & Selbst point to the risks of biased data and argue that data

mining can unintentionally reproduce historical discrimination resulting

from widespread prejudice in society.

256

In their conclusion, they

acknowledge the potential benefit to using “data mining to generate new

knowledge and improve decision making that serves the interests of both

decision makers and protected classes,” but caution that such adoption must

253

. Kahneman et al., Noise: How to Overcome, supra note 174, at 41.

254

. This is one of the advantages of the use of AI in employment decision-making.

Although human can ignore rules, machines cannot. See the study described in Uhlmann,

supra note 217.

255

. See ERIC SIEGEL, PREDICTIVE ANALYTICS: THE POWER TO PREDICT WHO WILL CLICK, BUY,

LIE, OR DIE (2d ed. 2016) (describing how predictive analytics are currently being used

by the government and business to identify preferences and risks and noting that the use

of data about groups that have been historically discriminated against can result in

discriminatory outcomes); VIKTOR MAYER-SCHÖNBERGER & KENNETH CUKIER, BIG DATA: A

REVOLUTION THAT WILL TRANSFORM HOW WE LIVE, WORK, AND THINK (2013) (discussing the

potential bias from the likelihood of errors in contained in big data); CATHY O’NEIL,

WEAPONS OF MATH DESTRUCTION: HOW BIG DATA INCREASES INEQUALITY AND THREATENS

DEMOCRACY (2015) (discussing potential risks of big data).

256

. Solon Barocas & Andrew D. Selbst, Big Data’s Disparate Impact, 104 CALIF. L. REV.

671, 677 (2016).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 333

be done with care.

257

In Kate Crawford’s article, the Hidden Biases of Big

Data,

258

she states: “Hidden biases in both the collection and analysis stages

present considerable risks, and are as important to the big-data equation as

the numbers themselves.”

259

Crawford points to a number of examples

where the use of big data resulted in inaccurate predictions. She suggests

that unquestioning reliance on these predictions would have resulted in the

wrongful allocation of resources.

260

She emphasizes, “In the near term, data

scientists should take a page from social scientists, who have a long history

of asking where the data they’re working with comes from, what methods

were used to gather and analyze it, and what cognitive biases they might

bring to its interpretation.”

261

She correctly suggests that looking to the

sources and purposes of data, rather than the resulting numbers, is key to

potentially addressing these types of issues. While data mined from the

internet or purchased from data brokers will reflect societal prejudices and

errors, most of the AI fixes suggested in this paper do not use these types of

data sets, but rather rely on data sourced from the company itself or the

industry. These scholars, however, have done a great service in bringing

these risks to the attention of those working in technology. As indicated in

the previous Part, since these articles have been published, not only have

recent developments in AI been shown to be very successful in reducing bias

and noise in human decision-making, new tools are now being developed to

both detect and remedy the potential for human biases to creep into data

used in machine decision-making. Although most of the suggested AI

techniques described above do not create significant risks of discrimination,

the following Subparts will highlight some recent advances in AI specifically

designed to mitigate the potential for discriminatory results.

A. “Garbage In, Garbage Out”

The primary risk of incorporating AI in employment decision-making

described by almost every scholar in this field is the potential for

discriminatory outcomes. As explained above, data mined from the internet,

social media and data brokers is likely to be error-prone and reflect societal

257

. Id. at 732.

258

. Kate Crawford, The Hidden Biases in Big Data, HARV. BUS. REV. (Apr. 1, 2013),

https://perma.cc/JC8Q-YT8K.

259

. Id.

260

. Id.

261

. Id.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 334

prejudices.

262

Gender and racial stereotyping in society have created an

internet which reflects these prejudices resulting in, for example, some

disturbing results based on the algorithms that run Google searches.

263

Data

from these sources should not be used to identify traits a company should

look for when hiring a new employee when the harm can be the exclusion

of an entire gender, race, or community.

264

This potential problem is known

as “garbage in, garbage out.”

265

An additional type of risk results from using data sets skewed in favor

of a gender or race.

266

For example, if you run an algorithm on an

organization’s data set seeking to identify common traits in the top

performers, and 80% of those top performers are male, the results will also

be skewed in favor of the male gender. This appears to be the reason why

Amazon had to scrap its in-house machine-learning algorithm to sort

through resumes to hire the best candidates.

267

A group of programmers in

262

. See, e.g., Barocas & Selbst, supra note 256 at 674 (stating that “an algorithm is

only as good as the data it works with” and can “reflect historic patterns of prejudice”);

Pauline T. Kim, Data-Driven Discrimination at Work, 58 WM. & MARY L. REV. 857, 921-22

(2017) [hereinafter Kim, Data-Driven Discrimination] (describing that algorithms “built

on inaccurate, biased, or unrepresentative data can produce outcomes biased along lines

of race, sex, or other protected characteristics”).

263

In one study, Harvard professor Latanya Sweeney looked at the Google

AdSense ads that came up during searches of names associated with white

babies (Geoffrey, Jill, Emma) and names associated with black babies

(DeShawn, Darnell, Jermaine). She found that ads containing the word “arrest”

were shown next to more than 80% of “black” name searches but fewer than

30% of “white” name searches. Sweeney worries that the ways Google’s

advertising technology perpetuates racial bias could undermine a black

person’s chances in a competition, whether it’s for an award, a date, or a job.

Nanette Byrnes, Artificial Intolerance, MIT TECH. REV. (Mar. 28, 2016),

https://perma.cc/HXR7-QTH8 (citing Latanya Sweeney, Discrimination in Online Ad

Delivery, ACM QUEUE, Mar. 2013, at 1, 10).

264

. Although there have been developments in correcting bias on internet-

generated data sets (of purchased data sets sourced from the internet), there is a long

way to go. One suggestion to bring attention to potential flaws in the data set is the use

of data sheets. Timnit Gebru et al., Datasheets for Datasets (Apr. 16, 2019) (working

paper), https://perma.cc/H75E-U3EM. For a sample data sheet, see Datasheet for

RecipeQA, RECIPEQA, https://perma.cc/HVF3-VK3G (archived May 16, 2019).

265

. This term is used, for example, by Pauline Kim in Big Data and Artificial

Intelligence: New Challenges for Workplace Equality, 57 U. LOUISVILLE L. REV. 313 (2019);

David Lehr & Paul Ohm, Playing with the Data: What Legal Scholars Should Learn About

Machine Learning, 51 U.C. DAVIS L. REV. 653, 656 (2017); Sullivan, supra note 13 at 2;

Barocas & Selbst, supra note 256 at 684-87.

266

. Barocas & Selbst, supra note 256 at 684-687.

267

. Jeffrey Dastin, Amazon Scraps Secret AI Recruiting Tool That Showed Bias

Against Women, REUTERS (Oct. 9, 2018), https://perma.cc/5UPB-NHLE.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 335

their Edinburgh office developed an algorithm to identify candidates by

teaching the program to recognize some 50,000 terms which had shown up

on past resumes.

268

The data set used to identify those terms consisted of

resumes submitted to it since 2014, the vast majority of which came from

male applicants.

269

Because the algorithm had been fed data from primarily

men it awarded higher scores to the resumes submitted by male applicants.

Even when the sex of applicants was not used as a factor, it still downgraded

resumes that includes the term “women’s” such as women’s tennis or

women’s chess.

270

The algorithm also taught itself to look for verbs more

commonly associated with male engineers, such as “executed.”

271

Not only

did the program weed out female applicants, it also failed to meet its goal of

designating the best candidates.

272

The results appeared to be almost

random, demonstrating that the program did not even work. It was not just

a discriminatory algorithm; it was a deficient one as it was unable to

produce the desired result. Although some have pointed to this example as

showing that AI is biased, it actually just reflects that it was a poor design

which failure was compounded by using biased unbalanced data which

unsurprisingly created biased results.

273

This exact type of danger can be

remedied by addressing these risks in the design of the algorithm and by

balancing the data and/or increasing the diversity of existing data points as

discussed in the next paragraph.

Organizations are actively working on methods for creating better data

sets for use in AI. IBM is one company that has published its work on

creating balanced data sets. “AI holds significant power to improve the way

we live and work, but only if AI systems are developed and trained

responsibly, and produce outcomes we trust. Making sure that the system is

trained on balanced data, and rid of biases is critical to achieving such trust”

write IBM fellows Aleksandra Mojsilovic and John R. Smith.

274

In an article

268

. Id.

269

. Id.

270

. Id.

271

. Id.

272

. Id.

273

. Although the study does not disclose the gender of the twelve engineers who

developed the program, it is likely that they were male as Amazon’s software engineers

are “overwhelmingly male.” Rachel Goodman, Why Amazon’s Automated Hiring Tool

Discriminated Against Women, ACLU (Oct. 12, 2018), https://perma.cc/FNL8-UNDD. The

need for a diverse set of programmers is discussed infra.

274

. Aleksandra Mojsilovic & John Smith, IBM to Release World’s Largest Annotation

Dataset for Studying Bias in Facial Analysis, IBM RES. BLOG (June 27, 2018),

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 336

by IBM developer, Kenneth Jensen, he explains how balancing works.

275

create a balanced data set, developers duplicate the results from the less

frequent category, called boosting. Developers also discard the results of the

more frequent category, called reduction. They can also combine boosting

and reduction to obtain more balanced results.

276

The goal is to reduce the

impact of a skewed data set.

277

Thus, if you are using a data set to look for

the traits of successful programmers where 80% of your programmers are

male, you would balance the data set to more evenly reflect both genders to

improve your outcome.

278

When your data sets contain little or no information about certain

groups of people, your algorithm will not accurately evaluate people who

belong to that group. In a paper presented at the Conference on Fairness,

Accountability, and Transparency, the authors were able to create a test to

measure the accuracy of commercial classification algorithms.

279

The test

was able to demonstrate how unbalanced data inputs (77% male and 83%

white) resulted in facial recognition software with a 1% error rate for

identifying light-skinned men, but a 20-34% error rate for darker-skinned

women.

280

According to Joy Buolamwini, author of the paper and founder of

Algorithmic Justice League:

https://perma.cc/CE9N-K36C.

275

. Kenneth Jensen, Use Balancing to Produce More Relevant Models and Data

Results, IBM DEVELOPER (Sept. 19, 2016), https://perma.cc/38B8-5T58. The author

points out that by using test data sets, you can determine whether business objectives

are being met and correct for unbalanced data (80% males in a sample, for example). Id.

While this article is specific to the use of the IBM SPSS Modeler, the methods used can be

applied to other analytics programs.

276

. How to Handle Imbalanced Classification Problems in Machine Learning?,

ANALYTICS VIDHYA (Mar. 17, 2017), https://perma.cc/6YSZ-D36Y.

277

. For example, MIT’s facial recognition AI, created by primarily white male

students, was unable to recognize black female faces because of the lack of diversity in

the data set. In order to remedy this effect, IBM agreed to release an annotated data set

of facial images that was balanced in terms of skin tone, gender and age. Malik Murison,

IBM Takes Steps to Tackle AI Bias, INTERNET BUS. (June 29, 2018), https://perma.cc/MX4D-

CHT2.

278

. Id.

279

. Joy Buolamwini & Timnit Gebru, Gender Shades: Intersectional Accuracy

Disparities in Commercial Gender Classification, 81 PROC. MACHINE LEARNING RES. 77 (2018).

280

. One of the authors, Joy Buolamwini, decided to investigate this issue after

noticing that an art installation using facial recognition did not work as well for darker

skinned people. She was able to show the bias in the data by feeding 1,200 images of

darker-skinned people and women into the programs to discover the actual bias. She

suggests that white male programmers did not notice this flaw because the program

worked quite well on white male faces. Although IBM says it is not a direct result of

Buolamwini’s research, it has employed a more balanced model with a half a million

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 337

If the training sets aren’t really that diverse, any face that deviates

too much from the established norm will be harder to

detect . . . Training sets don’t just materialize out of nowhere. We

actually can create them. So there’s an opportunity to create full-

spectrum training sets that reflect a richer portrait of humanity.

281

By being aware of the risks involved in using data sets which do not

represent society at large, and repeatedly testing for any potential bias,

biased results can be mitigated.

Another potential solution to reduce the potential harm of biased data

sets is to drastically increase the diversity of data points reviewed. This

solution has been incorporated into credit checks by reviewing more data

points than those traditionally measured.

282

For example, to help those with

little or no credit history, ZestFinance created an algorithm that considers

tens of thousands of pieces of information beyond what is used for a typical

credit score (which uses a limited number of data points).

283

The same could

be done with employment data. In addition, data analytics can be used to

search for discrimination in the outcomes themselves. Companies, such as

ZestFinance, frequently test the results that their automated processes

produce to discover any discriminatory results that could harm

applicants.

284

images of a wider range of facial images. However, Ruchir Puri, chief architect of IBM’s

Watson artificial-intelligence system, acknowledges that her research bought up some

very significant points. Larry Hardesty, Study Finds Gender and Skin-Type Bias in

Commercial Artificial Intelligence Systems, MIT NEWS (Feb. 11, 2018),

https://perma.cc/5EK5-C439.

281

. Jason Bloomberg, Bias Is AI’s Achilles Heel. Here’s How to Fix It, FORBES (Aug. 13,

2018), https://perma.cc/9SM2-L6DT.

282

. ZestFinance Introduces Machine Learning Platform to Underwrite Millennials

and Other Consumers with Limited Credit History, BUSINESSWIRE (Feb. 14, 2017)

[hereinafter ZestFinance Introduces], https://perma.cc/H3GM-7NEE.

Affirm and ZestFinance are both founded on the idea that by looking at tens of

thousands of data points, machine-learning programs can expand the number

of people deemed creditworthy. In other words, algorithms fed by so much

diverse data will be less prone to discrimination than traditional human-driven

lending based on a much more limited number of factors. Among the insights

discovered by ZestFinance’s algorithms: that income is not as good a predictor

of creditworthiness as the combination of income, spending, and the cost of

living in any given city.

Nanette Byrnes, Artificial Intolerance, MIT TECH. REV. (Mar. 28, 2016),

https://perma.cc/3LYT-2F4R.

283

. ZestFinance Introduces, supra note 282.

284

. Id.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 338

Accenture is one of the organizations that has created AI that can test

data sets for bias.

285

Known as the “Fairness Tool,” it can examine the data

set for sensitive variables, identify and remove any coordinated influence

that would result in an unfair outcome, evaluate false positives and

negatives, and display the impact that the fixes have on the model’s

accuracy.

286

MIT’s Computer Science and Artificial Intelligence Laboratory

has also created a method for detecting and mitigating bias resulting from

under-represented segments of society in training data for machine

learning.

287

Another issue mentioned by critics of the use of AI in employment

decisions is that biases in the programmers themselves may produce biased

results.

288

In addition to failing to notice that the facial recognition software

did not work well on black female faces because it was almost 99% accurate

on white male faces, programmers may, due to their unconscious biases,

measure the accuracy of an algorithms without considering races and

genders other than their own.

289

For example, programmers might chose

inappropriate “target variables” or “class labels.”

290

Because more men tend

285

. Rumman Chowdhury, Tackling the Challenge of Ethics in AI, ACCENTURE BLOG

(June 6, 2018), https://perma.cc/MBJ6-XJDP.

286

. Id.

287

. Alexander Amini et al., Uncovering and Mitigating Algorithmic Bias Through

Learned Latent Structure, PROC. 2019 AAAI/ACM CONF. ON ARTIFICIAL INTELLIGENCE, ETHICS,

AND SOC’Y 289 (Jan. 27-28, 2019), https://perma.cc/G7DE-6VMV.

288

. Jack Clark, Artificial Intelligence Has a ‘Sea of Dudes’ Problem, BLOOMBERG

(June 23, 2016), https://perma.cc/U4QA-4BTD.

289

. There are numerous examples where male software developers have made

assumptions resulting in the exclusion of other groups. For example, there are games

that do not provide female avatars. In almost all games, a male character is the default

and in other games although the male character is free, users have to pay for a female

avatar. Madeline Messer, I’m a 12-Year-Old Girl. Why Don’t the Characters in My Apps Look

Like Me?, WASH. POST (Mar. 4, 2015), https://perma.cc/Z39Q-5LEV. See also SARA

WACHTER-BOETTCHER, TECHNICALLY WRONG: SEXIST APPS, BIASED ALGORITHMS, AND OTHER

THREATS OF TOXIC TECH (2017) (discussing how the only female programmer in an app

development meeting had her ideas dismissed because the males were relying on their

impressions of their stay at home wives’ shopping activities, stating, “Oh, 51% of the

women can’t be tech-savvy” and insisting that the women cared only about shopping and

other leisure related activities and eventually launching a product that failed.) In another

example, Slack’s use of a brown hand for its “add to Slack” function would most likely

have not occurred to a white male programmer. Slack employs the “highest percentage

of female and black engineers of any tech company.” Tobias Hardy, How Slack Is Doing

Diversity and Inclusion Right, LAUNCHPAD, https://perma.cc/FUU4-NEHG (archived

May 20, 2019).

290

. Target variables are what the machine is looking for, and class labels are how

the data is classified. Because defining these target variables and class labels are

sometimes subject to the discretion of the programmers, the potential for unintentional

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 339

to be programmers, their own biases could affect the process.

291

Because

researchers have become more aware of these less obvious types of risks,

new processes are being developed to remedy these situations. Prejudicial

classifications, data errors, and incorrect or missing variables, for example,

can be audited for and eliminated in a number of ways.

292

Another solutions

is to make the programmers’ thinking process more transparent by

requiring them to document what went into their algorithm prior to

creating it.

293

Bruneis & Goodman suggest eight criteria that developers

would need to identify for review by others.

294

By considering the

implications of classifications prior to the creation of a program and

auditing the outcomes, bias can be detected and mitigated.

295

The best

solution to eliminate bias in programmers, however, is to hire a diverse

group to create the programs meant to provide fair and consistent

employment decisions. As Fei-Fei Li, Chief Scientist of Artificial Intelligence

and Machine Learning at Google and Co-Director of the Stanford Institute

for Human-Centered Artificial Intelligence said:

discrimination against “systemically disadvantaged protected classes” may occur.

Barocas & Selbst, supra note 256 at 677-80.

291

. Clark, supra note 288. At Google, for example, only 10% of the employees

working on machine intelligence are women. Tom Simonite, AI Is the Future—But Where

Are the Women?, WIRED (Aug. 17, 2018), https://perma.cc/VCB5-AS9N.

292

. See Andrea Romei & Salvatore Ruggieri, Discrimination Data Analysis: A Multi-

Disciplinary Bibliography, in DISCRIMINATION AND PRIVACY IN THE INFORMATION SOCIETY: DATA

MINING AND PROFILING IN LARGE DATABASES 109, 122 (Bart Custers et al. eds., 2013)

(acknowledging the growth in discrimination discovery and prevention in data analysis,

the chapter provides an “annotated bibliography of the literature on discrimination data

analysis”).

293

. Robert Brauneis & Ellen P. Goodman, Algorithmic Transparency for the Smart

City, 20 YALE J.L. & TECH. 103 (2018).

294

. These include: the predictive goals of the algorithm and the problem it is meant

to solve, the training data considered relevant to reach the predictive goals, the training

data excluded and the reasons for excluding it, the actual predictions of the algorithm as

opposed to its predictive goals, the analytical techniques used to discover patterns in the

data, other policy choices encoded in the algorithm besides data exclusion, validation

studies or audits of the algorithm after implementation, and a plain language explanation

of how the algorithm makes predictions. Henrik Chulu, Let Us End Algorithmic

Discrimination, MEDIUM (Aug. 3, 3018), https://perma.cc/TPN4-WPZZ (summarizing

Brauneis & Goodman, supra note 293).

295

. See, e.g., Marco Tulio Ribeiro et al., “Why Should I Trust You?”: Explaining the

Predictions of Any Classifier, PROC. 22ND ACM SIGKDD INT’L CONF. ON KNOWLEDGE DISCOVERY

& DATA MINING 1135, (Aug. 9, 2016), https://perma.cc/AXS3-FM7M (explaining a

software system that helps make sense of algorithmic decisions called LIME); Edward

Ma, Anchor Your Model Interpretations by Anchors, MEDIUM (Sept. 8, 2018),

https://perma.cc/7X49-ADDY (expanding on LIME with the use of anchors to explain

algorithmic decisions).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 340

If we don’t get women and people of color at the table—real

technologists doing the real work—we will bias systems . . . This is

the time to get women and diverse voices in so that we build it

properly, right? And it can be great. It’s going to be ubiquitous. It’s

going to be awesome. But we have to have people at the table.

296

Not only will hiring more women and URMs as programmers reduce the

potential for bias, it will have a significant effect by addressing concerns

prior to the creation of an algorithm.

B. “Black Box”

Another criticism of the use of AI in employment decisions is known as

the “black box” problem, which results from the difficulty in pin-pointing

why a machine produced a particular outcome.

297

The concern raised is that

if AI outcomes cannot be explained, they may contain unknown biases.

298

While some researchers are looking for ways to provide an understanding

of algorithmic outcomes without opening the black box,

299

others have

proposed various ways to audit for fairness and debias algorithms.

300

One

296

. Kate Brodock, Why We Desperately Need Women to Design AI, MEDIUM (Aug. 4,

2017), https://perma.cc/3S43-RHUX.

297

. See, e.g., FRANK PASQUALE, THE BLACK BOX SOCIETY 3 (2015) (describing how a

“black box” can mean “a system whose workings are mysterious; we can observe its

inputs and outputs, but we cannot tell how one becomes the other.”); Danielle Keats

Citron & Frank Pasquale, The Scored Society: Due Process for Automated Predictions, 89

WASH. L. REV. 1, 6 (2014) (discussing that the “black box” problem with AI is that the

algorithm may convert “inputs to outputs without revealing how it does so.”); Kim, Data-

Driven Discrimination, supra note 262 at 921-22 (“An algorithm may be a ‘black box’ that

sorts applicants or employees and predicts who is most promising, without specifying

what characteristics or qualities it is looking for.”).

298

. Citron & Pasquale, supra note 297.

299

. See Riccardo Guidotti et al., A Survey of Methods for Explaining Black Box Models,

51 ACM COMPUTING SURVEYS 1 (2018) (providing a survey of literature on addressing the

black box problem); Andrew D. Selbst & Solon Barocas, The Intuitive Appeal of

Explainable Machines, 87 FORDHAM L. REV. 1085, 1129 (requiring developers to document

the reasoning behind the choices made in developing a machine learning model); Sandra

Wachter et al., Counterfactual Explanations Without Opening the Black Box: Automated

Decisions and the GDPR, 32 HARV. J.L. & TECH. (2018) (explaining why a particular decision

was made so that the subject can make a change to obtain a different result in the future

is more valuable than opening the “black box”).

300

. Technological ways to debias algorithms can be accomplished pre-processing,

in-processing, and post-processing. For a collection of research on these methods, see

Abhishek Tiwari, Bias and Fairness in Machine Learning (July 4, 2017) (unpublished

manuscript), https://perma.cc/U3XE-XPGZ. See also Joshua A. Kroll et al., Accountable

Algorithms, 165 U. PA. L. REV. 633, 637 (2017) (describing “how technical tools for

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 341

example of a tool developed to address this challenge and provide greater

transparency into the algorithmic process is known as Quantitative Input

Influence.

301

This method can help identify the reason for an algorithmic

outcome by measuring and displaying the influence of inputs on outputs. In

other words, the more prominent an input, the greater impact it had on the

algorithmic decision. Methods such as these can provide an understanding

of why an outcome was produced without having to peer into the “black

box.”

302

In addition, AI can be used to prevent and detect bias in the algorithmic

outcomes.

303

The following are examples of some of the methods that have

been developed to deal with potential discriminatory results in machine

learning.

304

At the 2016 Neural Information Processing Systems (NIPS)

conference, researchers demonstrated their method known as hard de-

biasing for reviewing and removing gendered stereotypes resulting from

biased training data.

305

At the 2018 International Conference in Machine

Learning counterfactual fairness testing was also shown to be effective in

rooting out bias.

306

With counterfactual testing, instead of ignoring

verifying the correctness of computer systems can be used to ensure that appropriate

evidence exists for later oversight”); Pauline Kim, Auditing Algorithms for Discrimination,

166 U. PA. L. REV. ONLINE 189, 190-91 (2017) [hereinafter Kim, Auditing Algorithms]

(describing how auditing the outcomes of “decisionmaking [sic] algorithms similarly

offers a method of detecting when they may be biased against particular groups”).

301

. Anupam Datta et al., Algorithmic Transparency via Quantitative Input Influence,

2016 IEEE SYMP. ON SECURITY & PRIVACY, 598 (2016).

302

. Lehr & Ohm, supra note 265, at 710 (cautioning that such influence tests are

not available for every type of machine learning algorithm). For a survey of literature

reviewing explainability and transparency models, see Amina Adadi & Mohammed

Berrada, Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI),

6 IEEE ACCESS 52138 (2018) (summarizing literature on the growing field of explainable

AI to mitigate the black box problem and future research trajectories).

303

. See Rumman Chowdhury & Narendra Mulani, Auditing Algorithms for Bias,

HARV. BUS. REV. (Oct. 24, 2018), https://perma.cc/YX2H-7MWT (discussing a fairness tool

to audit outcomes developed by Accenture).

304

. For a full and very helpful description of the machine learning process, see Lehr

& Ohm, supra note 265, at 670-702.

305

. Tolga Bolukbasi et al., Man Is to Computer Programmer as Woman Is to

Homemaker?: Debiasing Word Embeddings, PROC. 30TH INT’L CONF. ON NEURAL INFO.

PROCESSING SYS. (Dec. 5-10, 2016), https://perma.cc/9A8W-JCD4. In this experiment,

gender stereotyping was learned when the algorithm was trained on Google News as a

data set. The programmers were able to “fix” the bias by removing the gendered

stereotypes such as associating receptionist with female but leaving the association

between queen and female essentially debiasing the algorithm. Id.

306

. The authors of a paper started with the premise that a decision is fair towards

an individual if it is the same in (a) the actual world and (b) a counterfactual world where

the individual belonged to a different demographic group. Matt J. Kusner et al.,

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 342

protected attributes, social biases are taken into consideration.

307

Essentially, two models are created: one including potentially biased factors

(such as gender and race) and one without. An outcome is considered fair if

it is the same in both models. In other words, the outcome would be the

same whether the individual was male or female or one race or another.

308

While this study examined fairness in law school admissions, it could

equally be applied in employment decisions. Another way of discovering

and mitigating the effects of biased data is by including additional data. This

is known as the text classifier model, where wrongful negative associations

are supplemented with nontoxic labels.

309

In this study, a text classifier was

built to identify toxic comments in Wikipedia Talk Pages but the program

would also flag nontoxic references, such as to the word “gay.” In order to

eliminate false positives, a number of non-toxic data points were entered

the training data set such as “I am gay” and “I am a gay person” to counteract

the many instances of toxic uses of the word “gay” in the data set and

preventing the flagging of these neutral statements as toxic language. The

study demonstrates the ability of tools to mitigate bias without reducing the

accuracy of the model’s results. This type of program could be used to

eliminate inappropriate wording in job ads and employee reviews.

310

Another method, adversarial debiasing, removes undesired biases in

training data while meeting a fairness test.

311

In other words, an algorithm

Counterfactual Fairness, PROC. 31ST INT’L CONF. ON NEURAL INFO. PROCESSING SYS. (Dec. 4-9

2017), https://perma.cc/HR4Z-NC3C.

307

. Id.

308

. Id.

309

. Lucas Dixon et al., Measuring and Mitigating Unintended Bias in Text

Classification, PROC. AAAI/ACM CONF. ON AI, ETHICS, AND SOC’Y 67 (Feb. 2-3, 2018),

https://perma.cc/8FWL-7RRU.

310

. Recently, Cygnet System came under fire for putting out a job ad for a tech

position indicting that the desired candidate would be “preferably Caucasian.” Gabrielle

Sorto, A Company Posted a Job Ad Seeking ‘Preferably Caucasian’ Candidates, CNN

(Apr. 30, 2019), https://perma.cc/NC3P-DFYE. By using algorithms to both examine

wording and correct inappropriate wording, candidates and employees can be treated

more fairly.

311

. Brain Hu Zhang et al., Mitigating Unwanted Biases with Adversarial Learning,

PROC. AAAI/ACM CONF. ON AI, ETHICS, AND SOC’Y 335 (Feb. 2-3, 2018),

https://perma.cc/4ETA-4CPN. In a greatly oversimplified explanation, using census data

(X), to predict income bracket (Y), the authors wanted to maximize the algorithm’s

ability to predict (Y) while minimizing the model’s ability to also predict a protected

attribute such as gender (Z). Because X contained unwanted biases reflecting Y, the

researchers sought to remove the generalizations about the protected attribute from the

data. Because they were able to accurately predict Y without predicting Z, they met the

accepted measurements of fairness. They were successful in training a model to be

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 343

could be designed to predict an output variable based on an input variable

without predicting a protected variable (such as gender or race). The ability

to predict the output variable without predicting the protected variable

would signal that the algorithm produces a fair result. This type of algorithm

could be used to test that the AI being used to predict required skills for a

particular position is fair for all genders and races.

Due to the attention that scholars have brought to this issue, many

companies are now working on these types of solutions.

312

IBM’s AI Fairness

360, for example, is an open source library of tools available online to people

and entities who want to identify and mitigate bias in their own machine

learning programs.

313

The tools include seventy fairness metrics and ten

“state-of-the-art” bias mitigation algorithms.

314

A number of other

organizations are actively working on methods to detect and mitigate

potential discriminatory results in machine decision-making, including

“demonstrably less biased” while still performing “extremely well” on predicting Y. Id.

312

. Kalev Leetaru, Can Algorithms Save Us From Bias?, FORBES (Jul. 31, 2016),

https://perma.cc/5M7C-YFMG.

313

. Jesus Rodriguez, What’s New in Deep Learning Research: Reducing Bias and

Discrimination in Machine Learning Models with AI Fairness 360, TOWARDS DATA SCI.

(Sept. 24, 2018), https://perma.cc/9SF7-3B9V. See also Cynthia Dwork & Deirdre K.

Mulligan, It’s Not Privacy, and It’s Not Fair, 66 STAN. L. REV. ONLINE 35 (2013) (describing

tools to test for classification bias).

314

. AI Fairness 360 Open Source Toolkit, IBM, https://perma.cc/59YJ-3ZEV

(archived May 20, 2019). Represented research includes work by Rachel K. E. Bellamy.

Rachel K. E. Bellamy et al., AI Fairness 360: An Extensible Toolkit for Detecting,

Understanding, and Mitigating Unwanted Algorithmic Bias (2018); Flavio P. Calmon et al.,

Optimizing Data Pre-Processing for Discrimination Prevention, PROC. 31ST CONF. ON NEURAL

INFO. PROCESSING SYS. (Dec. 4-9, 2017), https://perma.cc/87ET-8XXQ; Michael Feldman et

al., Certifying and Removing Disparate Impact, PROC. 21ST ACM SIGKDD INT’L CONF. ON

KNOWLEDGE DISCOVERY & DATA MINING (Aug. 10-13, 2015), https://perma.cc/3HC4-JJ3W;

Moritz Hardt et al., Equality of Opportunity in Supervised Learning, at 3315-28, PROC. 30TH

INT’L CONF. ON NEURAL INFO. PROCESSING SYS. (Dec. 5-10, 2016), https://perma.cc/W2AQ-

XWCN; Faisal Kamiran & Toon Calders, Data Preprocessing Techniques for Classification

Without Discrimination, 33 KNOWLEDGE INFO. SYS. 1 (2011); Faisal Kamiran et al., Decision

Theory for Discrimination-Aware Classification, PROC. 2012 IEEE 12TH INT’L CONF. ON DATA

MINING (Dec. 10, 2012), https://perma.cc/38Q4-YKNM; Toshihiro Kamishima et al.,

Fairness-Aware Classifier with Prejudice Remover Regularizer, at 35, JOINT EUR. CONF. ON

MACHINE LEARNING & KNOWLEDGE DISCOVERY IN DATABASES (Sept. 23-27, 2012),

https://perma.cc/6Q45-PFM3; Geoff Pleiss et al., On Fairness and Calibration, PROC. 31TH

INT’L CONF. ON NEURAL INFO. PROCESSING SYS. (Dec. 4-9, 2017), https://perma.cc/2ES9-ZNRJ;

Rich Zemel et al., Learning Fair Representations, at 28, PROC. 30TH INT’L CONF. ON MACHINE

LEARNING (June 16-21, 2013), https://perma.cc/KE4D-XHMH; Hu Zhang, supra note 311.

See also Emily James, Practical Steps to Addressing Algorithmic Bias, FLEXMR BLOG,

https://perma.cc/K8U2-R3CA (archived May 20, 2019) (explaining algorithmic bias and

potential tools for its mitigation). See also research on debiasing word embedding in

Bolukbasi, supra note 305.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 344

Facebook’s Fairness Flow,

315

Pymetrics’ open-source Audit AI Tool,

316

Accenture’s Toolkit, and Google’s What-if Tool.

317

Overall, a responsible AI

program to reduce bias in employment decisions will start with the careful

consideration of the design of the algorithms, the ongoing monitoring and

correcting of data, and the auditing of outcomes for potential discriminatory

results.

The focus on whether AI is biased is somewhat misplaced. The more

important question is: Does the use of AI result in less biased, more

consistent, and more accurate decisions than flawed human decision-

making?

318

No amount of mentoring in the tech industry is going to fix its

diversity problem. No amount of training is going to fix flawed human

decision-making. By focusing on the potential harms of AI, we miss

opportunities to make workplace safer and more equitable for all

employees.

319

According to Tim O’Reilly, who is an uncanny predictor of

trends in technology, “AI is not the machine from the future that is hostile to

315

. Dave Gershgorn, Facebook Says It Has a Tool to Detect Bias in Its Artificial

Intelligence, QUARTZ (May 3, 2018), https://perma.cc/4Q7P-MYUW.

316

. Khari Johnson, Pymetrics Open-Sources Audit AI, an Algorithm Bias Detection

Tool, VENTUREBEAT (May 31, 2018), https://perma.cc/SL9U-WUP6.

317

. The What-If Tool: Code-Free Probing of Machine Learning Models, GOOGLE AI

BLOG (Sept. 11, 2018), https://perma.cc/9H7J-W3CJ. For more recent developments that

may have been created since the publication of this article, see arXiv.com, the largest

open source database of scientific papers.

318

. Algorithms beat individuals about half the time. And they match individuals

about half time,” Kahneman said. “There are very few examples of people outperforming

algorithms in making predictive judgments. So when there’s the possibility of using an

algorithm, people should use it. We have the idea that it is very complicated to design an

algorithm. An algorithm is a rule. You can just construct rules.” Paul McCaffrey, Daniel

Kahneman: Four Keys to Better Decision Making, ENTERPRISING INVESTOR (June 8, 2018),

https://perma.cc/YYD7-ZUB3.

319

. While this paper focuses on gender, further research must be done on gender

in combination with racial bias. Sheryl Sandberg notes that “More companies prioritize

gender diversity than racial diversity, perhaps hoping that focusing on gender alone will

be sufficient to support all women . . . But women of color face bias both for being women

and for being people of color, and this double discrimination leads to a complex set of

constraints and barriers.” Courtney Connley, Why the Gender Pay Gap Still Exists 55 Years

After the Equal Pay Act Was Signed, CNBC (June 10, 2018), https://perma.cc/WBH7-

8EU9. For the types of issues that URMs face in the tech industry, see Erica Joy, The Other

Side of Diversity, MEDIUM (Nov. 4, 2014), https://perma.cc/9HBB-4LP5 (describing the

need for more research into “the psychological effects of being a minority in a mostly

homogeneous workplace for an extended period of time”); Aston Motes, Why Aston

Motes, Dropbox’s First Employee, Chose MIT Over Caltech, FAST COMPANY (Nov. 14, 2014),

https://perma.cc/5WRY-MUAY (explaining why Silicon Valley lacks diversity); Salvador

Rodriguez, Why Silicon Valley Is Failing Miserably At Diversity, And What Should Be Done

About It, I.B. TIMES (July 7, 2015), https://perma.cc/BRR8-MT3W (describing one

woman’s experience about the lack of diversity in Silicon Valley).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 345

human values and will put us all out of work. AI is the next step in the spread

and usefulness of knowledge, which is the true source of the wealth of

nations. We should not fear it. We should put it to work, intentionally and

thoughtfully, in ways that create more value for society than they disrupt. It

is already being used to enhance, not replace, human intelligence.”

320

IX.LEGAL CONCERNS IN INCORPORATING AI INTO YOUR D&I PLAN

It should be noted that many scholars have already addressed Title VII

in connection with using AI in employment decisions and concluded that

liability under current law is not likely.

321

Disparate treatment requires

“intent,” and a plaintiff likely cannot show that a machine intended to

discriminate. The consensus is that any claim of algorithmic discrimination

would fall under disparate impact theory due to a facially neutral practice

(using AI) but would most likely be excused as job related and consistent

with business necessity.

With a claim of disparate impact, courts would almost certainly use the

test set forth in Griggs, which requires that there be a disproportionately

negative effect on a statutorily protected group.

322

If the screening or testing

results in a more diverse workplace, there is no discrimination. If more

women and URMs are disproportionately screened out, the algorithm could

320

. Tim O’Reilly, What Will Our Lives Be Like as Cyborgs?, THE ATLANTIC (Oct. 27,

2017), https://perma.cc/8Q46-YTGL. In addition to technical solutions to mitigating AI

risks, more attention is now being paid to how human-centric AI can be developed. The

European Commission is in the process of creating ethical rules for the development and

use of AI. The goal is to create AI which is trustworthy and that supports human rights

rather than harms them. This human-centric approach recognizes the potential of AI to

improve society. The two main requirements are that AI serve both an “ethical purpose”

and be “technically robust” meaning that it should do what it purports to do. It is very

important for companies developing AI solutions to reduce bias and noise while

increasing diversity to keep these considerations in mind. EUROPEAN COMM’N, ETHICS

GUIDELINES FOR TRUSTWORTHY AI (2019).

321

. See Barocas & Selbst, supra note 256 (discussing how disparate treatment as a

cause of action would not work with an allegation of discriminatory AI due to the

requirement of “intent” and that a claim of disparate impact would likely be defeated

under the business necessity test); Kim, Data-Driven Discrimination, supra note 262, at

910-11 (arguing for a new theory of liability in response to classification bias); Lehr &

Ohm, supra note 265, at 666 (explaining Barocas & Selbt’s argument that Title VII is not

sufficient to address algorithmic discrimination); Sullivan, supra note 13 (concluding

current disparate treatment law would not fit with AI and disparate treatment claims

could be overcome with a showing of business necessity). Cf. Stephanie Bornstein,

Antidiscriminatory Algorithms, 70 ALA. L. REV. 519 (2018) (arguing that stereotyping

theory, a form of disparate treatment could result in a finding of discrimination by AI).

322

. Griggs, 401 U.S. at 431.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 346

be reviewed, and any bias detected and mitigated. The company must be

able to show that the algorithms used in the AI are valid for use and

accurately measure what they purport to measure.

323

Another concern

raised is that setting a goal of increasing diversity in your algorithm could

result in a reverse discrimination suit. This argument is not likely to

succeed, due to the valuable and quantifiable business reasons to employ a

diverse workforce.

324

AI helps organizations hire the best candidates.

Provided the organization uses AI in a responsible and intentional fashion,

they would meet this standard.

325

An additional concern that has been raised is that the auditing for bias

and/or correction of biased outcomes from a machine decision would

violate the holding in Ricci v. DeStefano.

326

In Ricci the court held that the

City’s decision to throw out the results of a test used to determine

promotions because of a fear of a disparate impact suit (no URMs scored

high enough to be promoted and only white men would have been

promoted) was a violation of Title VII and constituted disparate

treatment.

327

The analogy is that correcting a biased outcome by fixing the

algorithm is similar to throwing out a test because of biased results.

328

However, Pauline Kim suggests that any such auditing would be to

prospectively revise the algorithm to reduce bias thus there would be no

adverse employment action.

329

In addition, she argues that because Title VII

encourages “voluntary efforts to comply with nondiscrimination goals,”

auditing algorithms for bias would not run afoul of discrimination law.

330

Because the use of AI would be implemented to increase diversity and

323

. Matthew T. Bodie et al., The Law and Policy of People Analytics, 88 U. COLO. L.

REV. 961, 1023 (2017).

324

. Courts have found affirmative action plans as a legitimate nondiscriminatory

reason for an employment action permissible under Title VII countering a reverse

discrimination claim. Johnson v. Transp. Agency, 480 U.S. 616 (1987); United

Steelworkers of Am. v. Weber, 443 U.S. 193 (1979), as cited in Bodie et al., supra note 323

at 60.

325

. See Part VI for discussion on how AI can identify the best candidates.

326

. 557 U.S. 557 (2009).

327

. Id.

328

. Kroll, supra note 300, at 694-95 (arguing that any such auditing would violate

the holding in Ricci).

329

. Kim, Auditing Algorithms, supra note 300 at 197-202. See also Mark MacCarthy,

Standards of Fairness for Disparate Impact Assessment of Big Data Algorithms, 48

CUMBERLAND L. REV. 102, 125-29 (2017) (explaining how recent case law would not

prevent the “development, modification, or use of algorithms that have a lesser disparate

impact through aiming toward statistical parity or equal group error rates”).

330

. Kim, Auditing Algorithms, supra note 300, at 191.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 347

create more uniform employment decisions, rather than correct a specific

problem, it would not be analogous to throwing out the results of a test (as

in Ricci) to avoid a disparate impact suit.

331

Another benefit to using AI over subjective human decision-making is

that noise can be reduced or eliminated. This will help organizations avoid

liability for inconsistent employment decisions.

332

Claims of discrimination

based on noise occurs when multiple reasons are given for an employee’s

termination or employees in identical situations are treated differently.

333

This variability in decision-making can lead to a lawsuit for discrimination.

For example, when multiple reasons are given for an employee’s

termination, courts are likely to find that this was pretext for

discrimination.

334

When an employee in a protected class receives an

adverse employment action, but a similarly situated employee who is not in

the class does not, courts would likely find discrimination.

335

Using AI to

create consistency in employment decisions will avoid these types of

lawsuits.

331

. Id.

332

. See William T. Bielby, Minimizing Workplace Gender and Racial Bias, 29

CONTEMP. SOC. 120, 123-27 (2000) (citation omitted) (“[P]ersonnel systems whose

criteria for making decisions are arbitrary and subjective are highly vulnerable to bias

due to the influence of stereotypes—as, for example, when individual managers have a

great deal of discretion with little in the way of written guidelines or effective

oversight.”); Barbara F. Reskin, The Proximate Causes of Employment Discrimination, 29

CONTEMP. SOC. 319, 323-27 (2000) (arguing organizations can minimize stereotyping and

bias by using “objective, reliable, and timely information that is directly relevant to job

performance in personnel decisions”); Bornstein, Reckless Discrimination, supra note 90

at 1098 (suggesting that reducing the opportunity for subjective decision-making can be

effective in reducing the effects of stereotyping and implicit bias).

333

. I use the term “noise” to maintain consistent terminology in this article. As

explained, noise is unjustified inconsistency and variability in human decision-making.

Although this term has not yet been used in the context of employment discrimination

law, I anticipate it will become a rich vein of research in future law review articles after

Kahneman’s book NOISE comes out in 2020 or 2021.

334

. Velez v. Thermo King de Puerto Rico, 585 F.3d 441 (1st Cir. 2009). See also

Pierson v. Quad/Graphics Printing Corp., 749 F.3d 530 (6th Cir. 2014) (reversing

summary judgment because multiple conflicting reasons for the termination would

present the jurors with a triable issue of fact); Hitchcock v. Angel Corps, Inc., 718 F.3d

733 (7th Cir. 2013) (reversing summary judgement on grounds that the jury should be

allowed to determine whether the reason for termination was a pretext due to the four

conflicting reasons given for her termination).

335

. Int’l Bhd. of Teamsters v. United States, 431 U.S. 324 (1977) (setting forth the

standard for “disparate treatment” cases which turns on whether or not the Plaintiff is

treated differently than someone “similarly situated.”)

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 348

One particularly promising use of AI to avoid discrimination suits is in

the area of promotions. Under Title VII, an action may be brought when an

employee in a protected class is passed over in favor of a similarly or lesser

qualified employee who is not in a protected class.

336

Courts have held that

subjective promotion criteria can raise an inference of discrimination.

337

Because promotion criteria is seldom explicit in tech start-ups, for example,

there is a great deal of inconsistency in the criteria used for promotions.

338

Although unconscious biases may give rise to men being preferred for

promotion over women, as Part II supra indicates, the courts have been very

inconsistent in their treatment of gender discrimination based on

unconscious bias evidence. However, plaintiffs have been more successful

in cases where inconsistent treatment of employees due to the use of

subjective criteria for promotions can be demonstrated.

339

As discussed

earlier the rates of promotion for men and women in the tech field are

significantly different.

340

An algorithm can be used to evaluate employees

using objective criteria to avoid these types of lawsuits as well. As

mentioned earlier, AI can also be used to achieve more objectivity and

consistency in conducting interviews through the use of online structured

formats or chatbots. Another benefit is that a chatbot used to conduct

interviews will not ask illegal questions as some one in five humans do

336

. Under Title VII, to prove promotion discrimination, an employee who is

qualified, but in a protected class, must demonstrate that they were passed over for

promotion in favor of a person with similar or lesser qualifications. Reeves v. Sanderson

Plumbing Prods., Inc., 530 U.S. 133, 148 (2000); Jacobs v. N.C. Admin. Office of the Courts,

780 F.3d 562, 575 (4th Cir. 2015).

337

. Watson v. Fort Worth Bank & Trust, 487 U.S. 977 (1988) (ruling that Title VII

provides relief for denial of promotion due to subjective criteria); Garrett v. Hewlett-

Packard Co., 305 F.3d 1210, 1217 (10th Cir. 2002) (ruling that evidence of pretext may

be demonstrated with the use of subjective criteria); McCullough v. Real Foods, Inc., 140

F.3d 1123, 1129 (8th Cir. 1998) (“[S]ubjective criteria for promotions ‘are particularly

easy for an employer to invent in an effort to sabotage a plaintiff's prima facie case and

mask discrimination.’” (citing Lyoch v. Anheuser-Busch Cos., 139 F.3d 612, 615 (8th Cir.

1998))).

338

. Pawel Rzmkiewicz, Recruiting Methods for Startups: Balancing Objectivity &

Subjectivity for Tech Roles, MODERN RECRUITER (June 1, 2017), https://perma.cc/U3BJ-

K9JZ.

339

. Butler v. Home Depot, Inc., Nos. C-94- 4335 SI, C-95-2182 SI, 1997 WL 605754,

at *7 (N.D. Cal. Aug. 29, 1997) (sustaining certification of a sex discrimination class action

challenging hiring and promotion practices and quoting expert testimony explaining that

“[i]n the context of a male-dominated culture, relying on highly arbitrary assessments of

subjective hiring criteria allows stereotypes to influence hiring decisions”).

340

. McKinsey, supra note 50.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 349

during the interview process.

341

By incorporating AI to make consistent

decisions, the risks of variability in employment actions will be lessened.

342

Despite the numerous advantages from a legal perspective to using AI

to achieve more objectivity in employment decisions, there are a number of

additional legal issues that warrant further attention. The following briefly

recognizes several of these issues. Organizations must know where their

data is coming from and actively seek to examine and remedy bias and

variability in it. Those seeking to incorporate AI into their employment

decision-making should guard against data mined from the internet,

especially social media, and from data brokers, as it is likely to be biased and

error-prone.

343

If using internal data, examine it for balance, meaning that

the set is not predominantly one race or sex. The most important aspect to

reducing algorithmic bias is making a significant investment in clean data.

In addition, take care in designing the program. Avoid irrelevant

classifications, vague outcomes (such as a “good employee”) or permitting a

single or homogenous group of programmers to create algorithms.

344

Determine the desired outcome in advance and test any assumptions. In

other words, organizations should not assume that those with a college

degree will be better employees than those without. Auditing of outcomes

is also recommended to uncover any discriminatory results from the use of

341

. Questions about race, age, disability, marital status, number of children, place

of worship, are not permitted during the interview process. Chatbots are given specific

questions to ask resulting in the solicitation of identical categories of information from

each candidate (known as structured interviews) which has been shown to not only

eliminate bias, but also avoid illegal questions. Anthony Tattersall, Using Analytics to

Stamp Out Workplace Bias, LAUNCHPAD, https://perma.cc/5B22-MHG5 (archived Apr. 16,

2019).

342

. As Kahneman explains, noise can be an invisible problem. “When you have a

judgment and it is the noisy judgment, your judgment is determined largely by chance

factors. You don’t know what these factors are. You think you’re judging based on

substance, but in fact, that’s just not the case.” Matias, supra note 172. In order to prevent

errors, Kahneman suggests using an algorithm rather than a person. He notes, however,

that people “hate to be replaced by an algorithm.” Id.

343

. See Siegel, supra note 255; Chandler, supra note 212; Pasquale, supra note 297;

Mayer-Schönberger & Cukier, supra note 255; and O’Neil, supra note 255.

344

. Firms, such as Pymetrics, match skills discovered with certain positions, not

skills with the definition of a “good employee.” When the game determines someone’s

risk comfort, this information would be used to suggest appropriate positions. For

example, you would want someone who is more risk averse in your legal or accounting

department and less risk averse in your sales department. One of the benefits of games

is that, unlike personality exams, you cannot guess what answers the employee is

seeking. Chris Ip, To Find a Job, Play These Games, ENGADGET (May 4, 2018),

https://perma.cc/3YJB-T95N.

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 350

an AI program.

345

If someone wants to override a machine decision, the

reasons should be documented and only in exceptional circumstances.

Furthermore, creating an AI Council ensures the quality of your data,

examine classifications for legitimacy, and continually monitor outcomes

for bias.

346

When conducting employee surveys, employers should inform what the

surveys will be used for and make sure to note that questions regarding

race, gender, health, and veteran status are voluntary.

347

If the employer is

gathering information from emails and DMs used on company equipment,

counsel will need to examine state law and the Stored Communications Act

and the Electronic Communications Privacy Act to ensure that such

monitoring is permitted.

348

Anytime a business uses and maintains data,

there are privacy and security issues.

349

If any type of employee testing is

done, statutes such as the Americans with Disabilities Act must be

considered to make sure that the use of any analytics program, especially

games, does not factor these individuals out in a way that violates the law.

350

345

. See Kroll, supra note 300 and Kim, Auditing Algorithms, supra note 300 for

discussion on legality of auditing algorithms.

346

. For guidelines on the creation of trustworthy AI, see EUROPEAN COMM’N, supra

note 320.

347

. For examples of illegal interview questions see Illegal Interview Questions,

BETTERTEAM (Jan. 6, 2019), https://perma.cc/HM8W-89GB

348

. Society for Human Resource Management, State Surveillance and Monitoring

Laws, https://perma.cc/U6QH-FFHR (archived Apr. 16, 2019). See also Ifeoma Ajunwa,

Algorithms at Work: Productivity Monitoring Applications and Wearable Technology as

The New Data-Centric Research Agenda for Employment and Labor Law, 63 ST. LOUIS U. L.J.

(2019, forthcoming) (examining potential legal issues regarding productivity

monitoring of employees).

349

. See, e.g., Ajunwa, supra note 348; Ifeoma Ajunwa, Kate Crawford & Jason

Schultz, Limitless Worker Surveillance, 105 CALIF. L. REV. 735 (2017) (discussing worker

privacy issues due to advances in big data analytics, communications capture, mobile

device design, DNA testing, and biometrics); Ronald C. Brown, Measuring Worker

Performance Within the Limits of Employment Law in the Changing Workplace

Environment of Industry 4.0 (May 25, 2018) (unpublished manuscript),

https://perma.cc/6EGT-SRKY (discussing privacy and other legal consideration in using

technology to evaluate employee performance); Citron & Pasquale, supra note 297

(arguing that, because of the lack of legal oversight for automated decisions, due process

safeguards should be implemented); Pauline T. Kim & Erika Hanson, People Analytics and

the Regulation of Information Under the Fair Credit Reporting Act, 61 ST. LOUIS U. L.J. 17

(2016) (discussing limitations on employee data collection activities under the Fair

Credit Reporting Act); Karl M. Manheim & Lyric Kaplan, Artificial Intelligence: Risks to

Privacy and Democracy, 21 YALE J. L. & TECH. 106 (2019) (discussing risks of AI to

decisional and informational privacy and security).

350

. See, e.g., Allan G. King & Marko J. Mrkonich, “Big Data” and the Risk of

Employment Discrimination, 68 OKLA. L. REV. 555 (2016) (describing the potential for

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 351

If a discriminatory outcome is detected, rather than scrap the entire

program, like Amazon did, conduct further investigations as to the source of

the biased outcome and seek to remedy it. Research confirms that machines

are capable of not only making more accurate decisions than humans,

351

also confirms that in the area of employment decision-making, their ability

to override bias and noise will results in a greater diversity of hires, fairer

promotion decisions, and better retention of employees through early

detection of unhappiness.

X.CONCLUSION

Although the tech industry holds itself out as being committed to

diversity, it has failed to make any meaningful progress since the first

diversity report came out in 2014. It is estimated that U.S. companies will be

unable to fill the 1 million open tech positions in 2020. Because the tech

industry accounts for 20% of the country’s output, if the United States is to

remain an economic competitor in the world, it must be able to fill tech jobs

by vastly expanding its applicant pool beyond the usual suspects.

The tech industry and legal system have failed women miserably. The

excuses given for the lack of women in tech do not hold up. There are a

significant number of women from around the world who graduate with

computer science degrees and U.S. colleges have made advances in recent

years in increasing their numbers. However, many women are alienated

during the recruiting process due to gendered job ads and sexist behavior

during the interviews. The conditions women face in the tech field,

widespread disrespect, sexism, harassment, stereotyping, exclusion from

violating the Americans with Disabilities Act using data analytics).

351

. See, e.g., Wu Youyou, Michal Kosinski, & David Stillwell, Computer-Based

Personality Judgments Are More Accurate than Those Made by Humans, 112 PROC. NAT’L

ACAD. SCI. 1036, 1036 (Jan. 27, 2015) (“This study compares the accuracy of human and

computer-based personality judgments, using a sample of 86,220 volunteers who

completed a 100-item personality questionnaire. We show that (i) computer predictions

based on a generic digital footprint (Facebook Likes) are more accurate (r = 0.56) than

those made by the participants’ Facebook friends using a personality questionnaire (r =

0.49); (ii) computer models show higher interjudge agreement; and (iii) computer

personality judgments have higher external validity when predicting life outcomes such

as substance use, political attitudes, and physical health; for some outcomes, they even

outperform the self-rated personality scores.”); Nathan R. Kuncel et al., Mechanical

Versus Clinical Data Combination in Selection and Admissions Decisions: A Meta-Analysis,

98 J. APPLIED PSYCHOL. 1060 (2013) (stating that using mechanical means, such as

algorithms, to predict job or academic performance is 50% more accurate than holistic

methods, such as using experts or subjective human judgment).

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 352

networking events, and an inability to move up in the organization, cause

half of the women who enter the field to leave. In addition, courts have

shown a reluctance to find companies liable for non-overt discrimination

leaving a gaping hole in remedies. Class action suits are rarely certified, and

most tech companies have arbitration or confidentiality requirements that

prevent women from getting their day in court.

Although some believe that the tech industry’s homogenization is

intentional, it is more likely that the lack of diversity is due to the

unconscious biases and noise present in human decision-making. Because

most leaders in the tech field are men who rely on gut feeling and personal

opinion to make employment decisions, their unconscious biases have

created a vicious circle of hiring and promoting young white men. As Daniel

Kahneman explains, human judgments are untrustworthy due to cognitive

biases of which they are unaware. Mental shortcuts result in decisions

influenced by prejudice and bias. In addition, human decision-making is also

inconsistent. This noise presents itself as chance variability. Because these

flaws are not easy to counter in humans, the situation never improves. No

lawsuits, legislation or new theories of liability are going to solve this crisis.

This urgently needed solution must come from within the tech industry

itself.

It is only by introducing consistent objective decision-making into

talent-management decisions that bias and noise can be mitigated. This is

where AI is superior to humans. The use of AI in talent-management

decisions has shown success in not only creating more successful hires, but

in also creating a more diverse slate of candidates and employees. While

some companies have embraced these new technologies, others fear that AI

may actually cause discriminatory outcomes. As discussed, the phenomena

of “garbage in, garbage out” is real, but can be addressed paying attention to

the data sets by using known sources and making sure the sets are balanced

and representative of all groups. Additionally, the algorithmic process and

outcomes must also be monitored. The “black box” problem can be

addressed in multiple ways including testing for bias, providing indications

of influence, and auditing for fairness. By increasing the diversity of

programmers in the tech industry bias can be considered and prevented in

the creation of AI programs. As Fei-Fei Li warns, diverse programmers must

be hired now: “Trying to reverse [biased systems] a decade or two from now

will be so much more difficult, if not close to impossible.”

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 353

The growth of AI applications is not going to slow down, however, we

need to ensure that it is developed and used responsibly. With new

collaboration between disciplines such as law, psychology, economics,

business, engineering, technology and the social sciences, data sets are

being developed which more accurately reflect the demographics of the

society in which they exist and open source fixes are being created to

remedy potentially biased outcomes. By increasing diversity in the tech

industry, we will have more eyes and heterogenic perspectives overseeing

the development of AI. To be clear the answer is not to replace all human

decision-making with machines, but rather take advantage of the ability of

a machine to make decisions without noise (because an algorithm will

provide the same outcome for any given input, unlike the variability in

outcomes of human decision-making) and with less bias than humans

(because algorithms can be designed to review only the relevant

employment criteria unlike with human decisions).

Do the risks of incorporating AI into employment decisions outweigh

the benefits? From a purely legal standard point, it seems that despite claims

of increased discrimination using AI, scholars believe the risk of liability is

very small; however, the truth is more nuanced. Any potential

discrimination detected in outcomes will most likely stem from human

biases contained in data and not by virtue of the use of AI itself. AI alone will

not fix the diversity problem in tech but addressing the unconscious biases

prevalent in this industry is a first and vital step. It does not work to shame

management or require diversity training. Nor does it serve to delay

incorporating AI into your decision-making because of fear of

discriminatory results. What works is removing subjective criteria from

employment decisions, improving working conditions and the culture for

everyone in these tech companies, and providing oversight and

accountability for creating a diverse working environment. Most

importantly, while it is clear that bias cannot be removed from human

decision-making, it can be mitigated with machine decision-making. In fact,

with the rapid development of responsible AI, there may come a time in the

not so distant future when courts will find companies still using human

decision-making in employment to be a prima facie showing of

discrimination.

While this paper mentions specific solutions as examples of ways to

incorporate AI into employment decision-making, it is not meant as a

limitation, but rather a starting place. It is intended to serve as an alternative

Spring 2019 CAN AI SOLVE THE DIVERSITY PROBLEM? 354

more optimistic view in contrast to the dire warnings about the use of AI. AI

has enormous potential to address societal ills by quickly and efficiently

discovering where bias exists and how to root it out. This can have

enormous implications for correcting the societal injustices befalling

historically disadvantaged groups.

As Malcolm Gladwell wrote in TIPPING POINT, “Social change has always

followed this pattern; slowly emerging theories or paradigms accelerate

exponentially, suddenly appearing everywhere.” There is no stopping the

development of AI at this point, and there is no reason to. I encourage the

tech industry to take the lead in developing it responsibility—not only for

the benefit of their own organizations—but for the benefit of society and

our economy. We have jobs that need to be filled, and we know that human

decision-making is flawed. Addressing the underlying problem of noise and

bias in human decision-making will shift the needle towards a more diverse

and inclusive workplace. It is my hope that companies will make these fixes

open source, share best practices, and advance public understanding of how

AI can be used for the greater good. The question is not whether AI should

be incorporated into decisions regarding employment, but rather why in

2019 are we still relying on faulty human-decision making.