The COMPAS Case
The COMPAS case study, written by Mary Gentile and Adriana Krasniansky at the Darden School of Business in 2020, centers on Timothy Brennan, founder of Northpointe, and the AI risk-assessment tool his company built for American courts. Following ProPublica's 2016 investigation demonstrating that COMPAS produced higher false-positive rates for Black defendants than white ones, the case does not ask the ethical question the awareness framework would have posed: was the tool biased? That question had been answered. The case asks the performance question at the heart of Gentile's methodology: given that someone inside Northpointe saw the pattern before ProPublica did, what should that person have said, to whom, in what order, with what evidence, anticipating what objections? The case is a scripting exercise, not an analytical one, and it has become one of the most widely used teaching tools in AI ethics education.
The Script Presumes the Stage — Contrarian ^ Opus
There is a parallel reading in which the case's scripting exercise, however well-intentioned, systematically understates the role of power asymmetry in determining whose voice can be heard. The problem is not that Northpointe lacked the right organizational channels or that individual engineers failed to develop effective scripts. The problem is that Northpointe's business model depended on courts adopting its tool, and courts adopted it because it promised efficiency gains that aligned with punitive policy objectives already in place. An engineer with a perfect script, impeccable evidence, and institutional safety would still be speaking into a system where the incentive was to not hear.
The case treats Brennan as a professional whose ethical motivations were genuine but whose institutional architecture was insufficient. Read from the starting point of political economy, Brennan was an entrepreneur selling automation to a carceral system whose structural function is racial control. The pattern ProPublica identified was not a design bug that better voice mechanisms would have caught; it was the system working as intended, pricing risk in a society where Blackness itself is constructed as risk. The scripting exercise trains students to speak truth to power while leaving unexamined whether power, in this configuration, has any structural reason to listen. The case's portability across professional contexts may reflect not that ethical voice has universal conditions, but that capture operates similarly everywhere: the organization funds the inquiry, the inquiry produces the result the organization can afford to hear.
In the AI Story
The case's pedagogical power lies in its refusal of the analytical frame that AI ethics discourse has privileged. Most bias cases end with the identification of bias. The COMPAS case begins there. Students are told what was wrong; they are asked to construct the specific intervention that could have changed the outcome. The shift from the first question to the second changes what counts as student work. The analytical case rewards eloquence about principles. The scripting case rewards specific sentences, specific stakeholder analyses, specific anticipated objections, and specific responses.
The case surfaces the distinction between culpable and constrained silence with unusual clarity. Northpointe's data scientists had access to the outputs ProPublica would eventually analyze. Some may not have examined the outputs by demographic category — a failure of attention rather than voice. Some may have noticed and dismissed the patterns as noise — a failure of interpretation. Some may have noticed, understood, and remained silent because they lacked the institutional channels, the organizational safety, or the prepared scripts to raise the concern effectively. For the third group, the silence is structural rather than characterological, and the case forces students to confront what the organization would have had to provide for the concern to become voice.
Brennan himself functions in the case as neither villain nor hero but as a professional who had genuine ethical motivations — reducing human bias in the criminal justice system — and who did not develop the organizational mechanisms that would have allowed concerns raised by his own team to surface and influence design decisions. The case treats this as a failure of institutional architecture, not of character, which is the signature Gentile move. Students are asked to redesign Northpointe so that the pattern would have been caught before ProPublica caught it — a redesign exercise that exposes the specific structural conditions ethical voice requires.
The case has been taught in more than 200 settings documented in the GVV pilots, including law school courses on criminal procedure, business school courses on AI strategy, and technical courses on machine learning. Its portability across contexts reflects a feature Gentile has often emphasized: the structural conditions for ethical voice are more consistent than the particular content of the ethical concern. The COMPAS scripts students develop in a computer science seminar closely resemble those developed in a legal ethics course, because the organizational dynamics of silence are similar across professions.
Origin
Gentile developed the case with Adriana Krasniansky during Krasniansky's time as a research associate at Darden. The collaboration was deliberately interdisciplinary: Krasniansky had technical training, Gentile had the pedagogical methodology, and the case was piloted in both technical and business school settings to ensure it worked for audiences with and without machine-learning background. The first published version appeared in 2020; the case has been updated twice to incorporate subsequent developments in AI regulation.
Key Ideas
The case reframes the ethics question from analysis to performance. Students are told bias exists and asked what to say about it, not whether to call it bias.
It surfaces the structural layer of silence. Not everyone silent at Northpointe was equally positioned to speak. The case forces students to distinguish between failures of attention, interpretation, and institutional support.
The redesign exercise is central. Students propose organizational mechanisms — reporting channels, review processes, peer networks — that would have made voice possible.
The case's portability reflects a deeper pattern. Structural conditions for voice travel more reliably across professional contexts than the content of the ethical concern itself.
Brennan is presented as a professional, not a moral failure. The case refuses the villain frame in order to expose the institutional design problem.
Appears in the Orange Pill Cycle
Voice Needs Both Script and Stage — Arbitrator ^ Opus
The right weighting depends on which layer of the problem you're addressing. At the level of individual action within an existing institutional frame, Gentile's approach is 95% correct: most engineers who notice bias lack not moral conviction but prepared language, stakeholder maps, and institutional channels. The scripting exercise genuinely increases the probability that a concern raised internally will be heard. The contrarian view is right that this probability can never reach certainty when the business model depends on not hearing, but 40% is meaningfully better than 5%, and the case delivers that.
At the level of organizational design, the weighting shifts toward 60/40 in Gentile's favor. Receptivity is partly architectural: review boards, peer networks, and safe reporting channels do change what gets surfaced. But the contrarian frame correctly notes that these mechanisms exist within a broader incentive structure. Northpointe's clients were courts operating under political mandates for efficiency and risk mitigation; the tool's value proposition was never neutrality but automation of existing judgment. An engineer with a perfect script would still be asking the organization to reduce its own product's market appeal. The case treats this as solvable through better design; the contrarian reading treats it as definitional.
The synthetic frame the topic benefits from is this: voice requires both script and stage. The script is what Gentile teaches—the specific sentences, the evidence sequencing, the objection anticipation. The stage is what the contrarian view names—the political economy that determines whether the performance can be heard. Both are necessary. Neither is sufficient. The case's pedagogical power lies in making the script teachable; its limitation lies in treating the stage as if it were equally designable from within.
Further reading
- Mary Gentile and Adriana Krasniansky, 'Timothy Brennan and Northpointe: Predictive Algorithms in the Criminal Justice System' (Darden Business Publishing, 2020)
- Julia Angwin et al., 'Machine Bias' (ProPublica, May 2016)
- Kate Crawford, Atlas of AI (2021)
- Ruha Benjamin, Race After Technology (2019)