In my new paper, I explain how the creation of responsible artificial intelligence (AI) can address why women and under-represented minorities have a difficult time gaining a foothold in male-dominated industries. This is an especially important topic today as companies may hesitate to use AI because they fear that it may create discriminatory outcomes.
The ability of job applicants to submit materials online can leave employers with far more information than they need, giving rise to new technological methods for screening applicants. However, some companies have gone further by creating AI programs to not only screen resumes and cover letters, but to evaluate the applicants themselves through chatbots and gamified solutions. Scholars have noted, however, that these unregulated AI programs can lead to biased hiring results. This can occur due to bias in data sets, labeling, or the algorithm itself. In Big Data’s Disparate Impact, Barocas & Selbst explain how data mining from the internet can replicate prejudice in society (“garbage in, garbage out”). In the Hidden Biases of Big Data, Kate Crawford reveals how the inability to understand why AI produced certain results can lead to inaccurate outcomes that go unquestioned (the “black box problem”).
In my paper, I argue that you can create better results by using known sources for data and looking for and correcting bias in the data set. For example, if a data set is skewed in favor of one gender or one race, it can be balanced through boosting and reduction. Boosting would involve replicating the records of the lesser represented group while reducing would involve removing some of the records of the over-represented group, permitting the data set to be more reflective of society. Companies and universities are not only working on making balanced data sets available, they also have created AI to detect these problems in the data sets. Additionally, if a data set contains prejudicial information, that too can be weeded out with AI. A text classifier, for example, can add neutral terms to counteract toxic terms in a data set to prevent the false flagging of harmless language.
The black box problem is a bit more complex. The concern is that if you do not understand why an algorithm produced a certain result, you are unable to determine if that result is biased. Researchers have also found ways to counter this issue. One example is by implementing quantitative input influence analysis, which involves identifying the impact an input (category of data such as gender or race) has on an output (such as a choice to interview an applicant). Another is counterfactual testing, which involves creating two models: one that includes the suspect category (such as race or gender) and one without. If the outcome is the same using either model, it is considered fair. In other words, race or gender did not play a role in the result. While these solutions do not explain why an algorithm produced a certain result, they do serve to root out bias.
One of the side effects of introducing AI into employment decisions for either efficiency or accuracy is the ability to create a more diverse workforce, because objective decisions are shown to increase the hiring and promotion of women and under-represented minorities. Social scientists, such as Daniel Kahneman, have long explained that deficiencies in human decision-making result from both unconscious biases and noise. Humans use mental shortcuts to process information that unfortunately include irrelevant factors due to their unconscious biases. Additionally, newer research by Kahneman demonstrates how inconsistent human decisions can day-to-day and perhaps even hour-to-hour (which he terms noise) result in very subjective decisions. Because humans are unaware of their deficiencies in thinking, they are unable to make objective decisions. Even when informed of their unconscious biases, not only do human fail to correct their thought processes, unconscious bias training has been shown to actually increase bias.
These unconscious biases are especially problematic in employment decisions. With representative bias, an applicant who matches what a decision-maker first envisions when thinking about the job will probably be hired. This leads to cyclical discrimination. In other words, if a company’s computer programmers are 90 percent male, the decision-maker will unconsciously conclude that the ideal programmer will be male. With noise, humans are unable to recognize how inconsistent their own decisions are. Because the ability to treat everyone equally is the cornerstone to fairness, the inability of humans to make consistent decisions is especially harmful in hiring. These poor decisions are difficult to detect and root out because of the validity illusion, which allows humans to overrate their ability to make predictions. This is compounded by confirmation bias, which results in humans focusing on only the information that fits their pre-existing beliefs.
In my paper I conclude that, because human decision-making is so unreliable, the use of AI results in much more objective and accurate hiring decisions. However, AI is also much less biased when created responsibly. One of the most promising aspects of my research is that many of the companies and universities working in this area are making their solutions open source and available to anyone interested in better hiring. The use of artificial intelligence is only going to increase. Rather than decry the risks, companies should see AI as a potential solution to the pervasive problem of discrimination due to faulty human decision-making.
This post comes to us from Kimberly A. Houser, an assistant clinical professor of business law and the law of emerging technologies at the University of North Texas. It is based on her recent paper, “Can AI Solve the Diversity Problem in the Tech Industry? Mitigating Noise and Bias in Employment Decision-Making,” available here.