Economist Mackensie Alston, Ph.D, visited Swarthmore last week to deliver this year’s iteration of the annual Frank C. Pierson lecture, a talk entitled “Constructing the Criminal: Race and the Perception of AI-Generated Suspects.”
Alston is an assistant professor in the Department of Finance at the University of Illinois at Urbana-Champaign and a faculty affiliate for the Wilson Sheehan Lab for Economic Opportunities at the University of Notre Dame. Her most recent work utilizes AI-generated images, differentiated exclusively by race, to ascertain biases in professional police candidates and improve hiring practices.
The Pierson lecture, hosted by Swarthmore’s economics department, honors former department chair Frank Pierson ’34, a scholar whose interest in anti-discrimination work in labor economics aligns with Alston’s area of study.
Alston began by recalling the beginnings of her interest in the field of economics, highlighting psychology and business scholar Dan Ariely’s 2008 book “Predictably Irrational.” The text aims to dispel the notion that human agents are rational, common-sense actors — a core principle of traditional economic theory. Ariely points to a study which found that participants preferred $1 bags of Hershey’s Kisses to identical bags offered free of charge. This ostensibly inexplicable misalignment sparked Alston’s desire to understand the true background for people’s decision making.
Her interest originated from a desire to document and describe phenomena that confused her: “I see something when I look around in the real world and there’s not a theory to explain it … [I] design experiments so that I can actually document this scientifically and prove that this phenomenon exists.”
Alston elaborated on the process of scientific exploration that suits pursuing answers to economic questions. For economists, experiments come in various forms: lab experiments, which operate through controlled, artificial environments; online experiments, which allow for speed but lack environmental control; and field experiments, which often feature unknown participants and enable realism but provide researchers with even less control. Additionally, economists frequently make use of observational data, for which they lack any ability to control or manipulate circumstances or randomization of treatment.
Alston noted that her favorite definition for economics is as “a study of decision-making” because it relates most to the field’s highly experimental nature. In turn, she emphasized the positive implications of conclusive research. Identification of a legitimate and salient causal relationship through study enables policymakers to prove the efficacy of ideas before they are officially widely implemented, as well as to compare theoretical predictions with real world results.
The criminal justice system and economics
In 1968, Gary Becker theorized the rational criminal model, a principle of incentive-weighing which has since been applied to the real-world criminal justice system. The model accounts for various factors influencing the decision to commit a crime, such as probability of punishment, severity of punishment, and payoff from criminal and noncriminal options. It’s evolved to include factors such as attitude towards illegal and legal behavior as the field more broadly has grown to comprise a variety of actors beyond the criminal themself. Alston’s work seeks to examine the factors behind a police officer’s decision to use potentially lethal force and how to identify racial biases that might influence that decision.
Alston and her partner, Dr. Emily Owens, were inspired by a breadth of economic research showing that Black people are mistreated by the police in a variety of ways. Further research has explored methods to combat this prejudice; in particular, improved officer anti-bias training has been shown to have positive impact. However, anti-bias training faces limitations in efficacy due to its mandatory nature and a tendency for its effects to dwindle over time, as well as the required continuous support of supervisors necessary for long term effectiveness. Another method, which Alston’s research focuses on, has to do with improving bias in new police officers rather than improving the bias of already hired police officers.
Alston also cited previous literature, including a 2017 study that identified disparities in Black NFL players being judged to be heavier, taller and more capable of harm than their white peers. Through utilization of AI, Alston’s work seeks to formulate an even more exact isolation of race as the only divergent factor between the images people judge so as to pinpoint participant racial bias.
She has, so far, conducted two studies making use of AI images, created using Midjourney and Gemini Proshop. Additionally, she plans to iterate on her previous work through further study, in part by making use of the feedback she receives at lectures such as this one.
For a more fully realized picture of action-oriented racial bias, Alston drew on a 2002 study which used real life images of people overlaid over background location images. The study then had participants, who act as police in the game, shoot the armed figures and not the unarmed figures, assigning points for correct decisions and penalties for incorrect decisions. The results of both studies themselves, though shared at the campus event, haven’t yet been published and made public due to the intention of conducting future studies.
After the lecture, Alston gave those in the audience a chance to ask questions. One student wondered about the effects of certain hair styles on racial bias, which Alston highlighted as a valid concern, noting that a Black person with a given hairstyle might be perceived differently than a white person with the same hairstyle.
Another audience member asked how the social desirability bias might be resolved in future studies. Alston replied that she’s considering asking participants to describe how others might describe the images, rather than describing the images directly themselves. Although it wouldn’t indicate how the participants themselves were biased, it would inform broader understandings of the perception of the images across participant groups.
Another student questioned whether the fact that the images were AI-generated might have shaped the biases people demonstrated towards them. Alston clarified that they had asked if participants thought the stimuli resembled AI, and received about a 70/30% split leaning towards AI.
In a brief interview with The Phoenix after her lecture, Alston emphasized the value of interdisciplinary analysis. For example, while economists weigh cost-benefit analyses and incentivize positive behavior, sociologists are able to provide cultural and behavioral context, and psychologists can provide insight into individuals’ psyches. In the midst of cross discipline contribution, the academic community as a whole can begin to approach the right solution.
