U.S. Equal Employment Opportunity Commission
U.S. Senate Equal Employment Opportunity Commission – Written Testimony of Michal Kosinski, Assistant Professor Organizational Behavior Stanford GSB
Meeting of 10-13-16 Public Meeting on Big Data in the Workplace
Madam Chair and Commissioners, thank you for the opportunity to testify today about the use of Big Data in recruitment and its influence on employment discrimination. I am Michal Kosinski, Professor of Organizational Behavior at the Stanford University Graduate School of Business. My research focuses on individual differences in behavior, preferences, and performance. Specifically, I am interested in the mechanisms linking psychological traits (such as ability or personality) with a broad range of organizational outcomes, including job performance and person-job fit. Additionally, I am studying the applicability of artificial intelligence and Big Data mining in occupational assessment.
In the following testimony, I will argue that if used properly, Big Data-coupled with modern computational techniques-can improve person-job fit, increase our ability to identify talent, raise equality in access to jobs and careers, and help overcome implicit and explicit prejudice in the workplace. This provides an opportunity to not only reduce discrimination (the focus of this testimony), but also boost the competitiveness and productivity of our economy, and the well-being of employees.
Importantly, Big Data models aimed at recruitment and performance appraisal are likely to disproportionally benefit groups that are currently the most discriminated against. White males, equipped with strong social networks and diplomas from leading universities, are already doing well in the job market. People that are likely to gain the most from the methods described below include women, members of racial minorities, those who lack access to education and mentors / role models, and people who are short of cultural and social capital.
Predicting Real-Life Outcomes with Big Datasets of Digital Footprints
We are increasingly surrounded by digital products and services that mediate our activities, communication, and social interactions. Consequently, a growing fraction of human behavior, thoughts, and feelings leave digital footprints. Such footprints include web browsing logs, transactions from online and offline marketplaces, photos and videos, GPS location logs, media playlists, voice and video call logs, language used in Tweets or emails, and much more. The amount of such data is staggering. Back in 2012, IBM estimated that people produce 2.5 billion gigabytes1 of data every single day, or about 350 megabytes for every inhabitant of Earth. Much of this data is recorded and archived, producing big datasets of digital footprints.
A quickly growing body of research shows that big datasets of digital footprints can be used to accurately predict a broad range of real-life outcomes and psychological traits. Such outcomes include career choices, intelligence, personality, happiness, and many other variables relevant in the occupational context. Such predictions are also remarkably accurate. It has been shown, for instance, that a prediction model based on Facebook Likes (a relatively generic type of digital footprint) is more accurate at predicting an individual’s personality than their spouse.2
In my own research exploring many different varieties of digital footprints, I have repeatedly observed that, given a training dataset that is large enough, computer models can be trained to predict virtually any outcome. This is because behaviors, preferences, and real-life outcomes are rarely random; instead, they form intricate patterns. Some of these patterns are fairly obvious and can be spotted by humans or traditional assessment tools, as is well illustrated by currently available recruitment and career planning tools assessing future job performance. Most of such patterns, however, are much weaker and can only be detected by computer algorithms in huge datasets. When applied to individuals, such algorithms can combine huge amounts of weak pieces of evidence to produce a very accurate prediction.3
How to Properly Use Digital-Footprint-Based Models in Recruitment and Appraising Performance
As I will explain in the next section, the digital-footprint-based models can help in reducing discrimination. To do so, however, they need to be used properly. Fortunately, decades of research in psychometrics and industrial/organizational psychology produced a set of principles aimed at conducting occupational assessment and performance appraisals in an accurate, valid, and unbiased fashion.
While digital-footprint-based tools significantly differ from traditional questionnaire or interview-based assessment tools, such principles of assessment still need to be applied. One of such principles, often overlooked in the discussion of digital-footprint-based models, states that only the factors causally linked with a performance in a given job can be employed in the ranking of the candidates. While hiring software engineers, for instance, one can rank them based on their conscientiousness or cognitive abilities, but not based on their gender.
The same principle has to be applied to digital-footprint-based models. As such models are usually so complex that it is impossible to fully comprehend their functioning, special steps have to be taken to ascertain that they do not discriminate against candidates based on the factors unrelated to job performance. Models that directly predict performance in a given job (or its correlates such as promotions, salary, or “being similar to John, our best employee”) are prone to recreating biases present in the training data. To avoid this issue and retain full control over the factors used in ranking the candidates, the models should not be aimed at job performance directly, but at the well-defined factors causally linked with job performance, such as personality, ability, or skills. The estimates produced by such models should then be checked for the evidence of any bias (the science of occupational assessment provides many tools that can be used to do so) and, finally, used to rank the candidates.
Reducing Discrimination in Employment Using Big-Data-Based Models
The following subsections briefly introduce the main applications of Big Data and computational methods in the context of employment, and discuss how they can help in combating discrimination.
Bias in recruitment is one of the main sources of discrimination in the workplace. Traditional methods used in recruitment, such as non-structured interviews and rating resumes, were shown to be poor predictors of a candidate’s performance and heavily affected by a number of biases. “Similar-to-me bias,” for example, drives a discrimination against groups underrepresented in a given profession. Recruiters’ decisions were also shown to be affected by a candidate’s ethnicity, name, and gender. Moreover, even somewhat widely accepted proxies for a candidate’s potential, such as the rank of their university, might lead to discrimination (in this case, by perpetuating the discrimination in access to the elite schools.)
Many of such biases could be avoided if unreliable and biased tools, such as non-structured interviews and resume-based rankings, were replaced with objective psychometric tools. This would not only reduce discrimination, but also improve the efficiency of the workforce, as people would be assigned to jobs based on their potential and not their gender, race, or name. However, objective psychometric tools, such as structured interviews and standardized ability tests, require expertise, are expensive to purchase or develop, and are time-consuming to administer. Thus, they are rarely used when recruiting for entry-level jobs, and are virtually absent from small organizations. As a result, a large part of the workforce (and, arguably, the part that is at the most risk of discrimination) is deprived of the benefits of objective assessment.
Digital-footprint-based models provide an inexpensive and powerful alternative to traditional occupational assessment methods. While the cost required to develop a traditional psychometric measure can reach millions of dollars, computational models can be developed quickly and at a fraction of this cost. Once developed, the marginal cost of assessing an individual is comparable with the marginal cost of conducting a Google search-and is as fast. By decreasing the cost of assessment, digital-footprint-based models could enable any candidate to be judged based on their objectively measured potential and not on a subjective and unavoidably biased opinion of a recruiter.
Performance appraisals constitute another significant source of discrimination in the workplace. Workers’ appraisals are often based on the subjective opinions of their peers or bosses, which are imperfect approximations of their actual performance. These appraisals are also heavily swayed by workers’ likability, similarity to the raters, and stereotypes associated with their gender or ethnic group. This problem is especially pronounced in professions where an individual’s performance is difficult to measure in an objective way, such as in law, consulting, engineering, and other highly specialized professions. (Such professions happen to be especially affected by discrimination.) The analysis of digital footprints generated in an occupational context provides an alternative to traditional performance appraisals provided by peers or managers. As I mentioned before, such models should not be trained on past peer-based ratings or aimed at recreating them, as this would most likely perpetuate the discrimination. Instead, digital-footprint-based models can be used to estimate objective performance indicators, such as the quality of the written texts, the amount and quality of interactions with other team members, etc.
Identifying talent rather than skills
Factors such as an individual’s socio-economic status and geographical location create large inequalities in access to education and vocational training. Thus, skills and knowledge are unevenly distributed across genders, geographical areas, and ethnic groups, translating into self-perpetuating barriers to jobs and careers. In other words, groups characterized by the same average ability (i.e. talent or potential) often significantly differ in the level of skills and knowledge, leading to unequal occupational opportunities.
Models based on digital footprints can be used to reveal latent psychological traits and future behavior. This is similar to what can be obtained using traditional ability and personality tests but, as discussed before, could be achieved much more inexpensively. In other words, such models can measure an individual’s ability and potential rather than their skills and knowledge. This is important, as in a typical work-sample test or an interview, even a mediocre candidate with some training can easily outperform a highly talented candidate deprived of training. Equipped with cheap and accessible tools to measure talent (rather than skills), the employers (and educational institutions) could identify people with a potential to excel at a given task given additional training. This would not only boost the equality in access to jobs, but also the workforce’s performance.
Cultural norms, stereotypes, and the lack of role models discourage members of some groups from seeking employment or training opportunities in certain professions. This problem could be confronted by using behavioral and demographic targeting tools that are available across many online platforms such as Facebook and Google. Employers and educational institutions can use such platforms to reach out to underrepresented groups with offers of training or employment. Moreover, this approach could be expanded by employing the passive assessment approach to make such offers more specific. Passive assessment employs the digital-footprint-based models to identify digital footprints that best predict a given characteristic (e.g. high conscientiousness). Knowing what digital footprints are typical of a person likely to obtain a particular characteristic enables the identification of such people in large, anonymous populations. For example, if visiting a particular set of websites has been shown to be linked with a talent for computer programming, one can target visitors of these websites with advertisements encouraging them to sign up for a programming course, or to apply for a given position.
Big Data Analytics
Finally, mining big datasets of digital footprints can provide policymakers with broad and instant information about the state of the job market, the distribution of skills, and potential sources or outcomes of discrimination. For example, participation in online courses, visited websites, and questions asked and answered in online forums can be used to track the distribution of skills and knowledge across demographic groups. This could provide policymakers with cheap and instant access to these variables, or track the performance of the interventions aimed at improving access to education and vocational training.
Many large databases are publicly available or could be easily obtained from various platforms. Excellent data sources that could provide meaningful insights related to discrimination in the workplace include Tweets, search trends, or Facebook Audience Insights.4
As outlined in this brief testimony, digital-footprint-based models offer the potential to reduce discrimination in the workplace. However, like many other new technologies, such models are not without risks. If used improperly or with malicious intent, they can perpetuate and amplify discrimination, rather than reduce it. Additionally, as I discuss more broadly in 2013 paper in Proceedings of National Academy of Sciences,5 such models pose particularly severe risks to privacy, as they could be also used to infer intimate traits such as sexual orientation, religion, or political views.
Given the great opportunities offered by Big Data, I would like to use this opportunity to make one important recommendation. I would like to encourage the Commission to facilitate the development and publication of a set of guidelines for the development of digital-footprint-based assessment models that are accurate, valid, and free from bias. This would encourage the development and adoption of computational models, and could have a strong and positive impact on reducing discrimination in the workplace. I would be thrilled to be a part of such initiative.
1 The stack of paper necessary to print one day’s worth of data produced by humanity in 2012 would need to be about 400 million kilometres tall-nearly three times the distance from the Earth to the Sun. (When using a regular A4 printer paper, covered on both sides with zeros and ones printed in 12pt-size font.)
3 A demo of a digital-footprint-based model (employing Facebook Likes) can be accessed at: http://www.applymagicsauce.com.
5 “Private traits and attributes are predictable from digital records of human behavior,” by M. Kosinski, D. Stillwell, T. Graepel, Proceedings of the National Academy of Sciences, 2013.