ADL new experimental study on discrimination against Jewish job candidates

The Antidefamation league released this week a new study that uses the longstanding design of identical resumes varying only names and signals of identity - such as participation in ethnic sports leagues or associations. The experiment includes a good number of observations, 3K emails: across many the United States the results, statistically significant and quite robust, were that Jewish sounding names got fewer callbacks than Western European names, and Israeli sounding names received even fewer.

A few observations: this was specifically for administrative jobs – would be interesting to replicate in other sectors and types of jobs  and specifically in lower skilled versus higher skilled jobs like high tech / programming / engineering / medicine / law. Second, all the (fictional) applicants were female – one could hypothesize that discrimination may be even more pronounced for Israeli male applicants given the past year and how protests have become violent in certain places, like in Los Angeles where the study seems to show some of the bigger gaps between applications. It is interesting that in only two cities – NY and Philly – there was no such finding of discrimination – which may be consistent with a high percentage demographically of Jewish employers.

Finally, I will underscore that indeed it is difficult to detect discrimination at the hiring stage [as opposed to firing/promotion when the employee has more info) so these studies are valuable; both federal and state laws protect against discrimination on basis of race, religion and ethnicity but discrimination in employment remains pervasive. I actually wrote about this longstanding method of resume studies in my book The Equality Machine - here is a snippet of that section:

Would an Algorithm Hire Lakisha Washington?

We worry that algorithms are black boxes—in other words, opaque and difficult to understand (which they often are). But what about the black box of the human mind? Human decision-making in the hiring realm involves dozens of recruiters, interviewers, co-workers, clients, and supervisors, each a small black box of their own. By contrast, using technology, we can check our intuition and innate human bias by employing machines to help us quantify and analyze information. We need to strive to integrate the best of both worlds—human and machine decision-making.

Two decades ago, a group of psychologists began running résumé experiments. They sent more than 5,000 identical fictitious résumés differing only in the applicants’ names to 1,300 employers in response to job ads posted in Boston and Chicago newspapers. The pretend applicants were named Greg Baker, Jamal Jones, Emily Walsh, and Lakisha Washington. The results were telling: “white-sounding” names received 50 percent more callbacks for interviews. That study was so illuminating that researchers all over the world began replicating it, manipulating other protected identities in the fictitious résumés. These studies have consistently found gender, race, age, and sexual orientation discrimination in hiring using résumé manipulation.

Twenty years of these résumé studies have been frustratingly consistent: despite social efforts and legal rules, human bias thrives. In all sectors—not least of all in the tech industry itself—despite decades of anti-discrimination laws on the books and diversity and inclusion training in place, workplaces still demonstrate bias in recruiting and hiring. To be sure, using technology to supplement or replace human decision-making carries risk and is not a panacea, but it has the potential to mitigate our innate human bias. University of Chicago professor Sendhil Mullainathan, who co-authored the original résumé study twenty years ago, argues that algorithmic bias is more readily discovered and more easily fixed than human bias. Studying what algorithms do, Mullainathan says, is “technical and rote, requiring neither stealth nor resourcefulness,” which makes discovering algorithmic discrimination more straightforward. Humans on the other hand, Mullainathan warns, are inscrutable in a way that algorithms are not. Even when the algorithms’ workings are opaque – or a blackbox – we can more systematically check the outcomes they produce to monitor for bias. When Mullainathan and his collaborators first conducted their résumé experiment—before the internet became the primary vehicle for job searching—it was a complex covert operation. They created banks of fictitious résumés, collected job opening data, faxed fake applications to prospective employers, and waited to receive job interviews or offers in order to identify the human bias that the study revealed. Nowadays, we can detect bias and imbalance in searches and screening in a much easier and more immediate way.

Technology also changes the way we can prove discrimination when disparity is detected. In my work as an expert witness in discrimination cases, I see how difficult it is to convince a judge and a jury that what happened to an employee was the result of bias. These cases have become even more difficult to prove as discrimination has become more subtle and furtive. Before Congress enacted Title VII of the Civil Rights Act in 1964, ads explicitly stating that women and minorities “need not apply” were commonplace in the job market. Now, the smoking gun of discrimination—such as the Idaho law specifying that “males must be preferred to females” in appointments for certain positions, a law that led to the landmark U.S. Supreme Court decision in Reed v. Reed—is mostly a thing of the past. Discrimination today is more subtle and more disguised. In hiring decisions, employers usually do not have formal, discernible rules on what weighs heavier among the many factors considered—experience, skill, education, personability, references, the likelihood that an applicant will accept an offer, and so on. Often, companies will just say that they are looking for the employee who is “the best fit.” Employment discrimination litigation is therefore notoriously difficult, especially when an applicant has not previously worked for an employer. And even when an employee has worked at the organization for a while, most evidence is circumstantial. Employers shift their explanations and proffer decision-making rationales that can be impenetrable to outside scrutiny.

Even more importantly, when we find that people are biased, what can we do about it? Litigation is a long, arduous, and after-the-fact process. It can financially compensate the employee who was discriminated against, but to what extent does it change hearts and minds—and most importantly, institutions? We can bring in sensitivity training and develop departments dedicated to diversity and inclusion, but it’s very hard to debias humans. Systemic, lasting change has been elusive.

Enter algorithmic decision-making. Done right, it can overcome the flaws of human decision-making. As Mullainathan says, “software on computers can be updated; the ‘wetware’ in our brains has so far proven much less pliable.” With these new pliant machines, we can expand how job opportunities are communicated; expand the applicant pool by identifying more inclusive formats and language; and employ screening measures that reject past, demonstrated human biases. We can then monitor and detect exclusions and continue to improve screening measures. As we explore each of these stages of the employment process in the following pages, we will see how, while a data point that an algorithm provides may be tainted by human bias and unequal realities, AI can continuously improve; algorithmic processes can be audited and corrected swiftly in a way that a human mind simply cannot. This malleability and adaptability vastly outclasses our current hiring practices, which rely on biases that continue to shape recruiting, mentoring, hiring, evaluation, and promotion processes.