Mehra- Blog Post 3

Johnson’s essay asserts that inductive reasoning entails making certain assumptions, or canons. Deciding which of these canons to adopt-- epistemic vs. non-epistemic, for example-- entails being allied to certain values over others. Further, inductive risk necessitates adopting a “threshold for confidence [that] can only be established by appeal to ethical values” (Johnson 13). These examples serve as objections to the value-free ideal, allowing Johnson to ultimately contend that “the extent to which many of the real-world applications of machine learning today are useful to the extent that they are undeniably value-laden” (Johnson 20). 

I’m especially intrigued by the AIPIE model that Johnson outlines in Section 4, and how the interaction between humans and algorithms may lead to more or less biased results. Johnson outlines the five-step AIPIE model as “Assessment, Interpretation of results, Plan based on the information gathered, Implement the plan, and Evaluate the results of the action” (Johnson 16). She further notes that in this model, an algorithm-based program like COMPAS would be employed in the “Assessment” phase, while practitioners may use the results of that phase in their interpretation, planning, and implementation. Regarding these practitioners, Johnson writes “rely[ing] on the judgments made by individual judges (unaided by risk-assessment tools) would likely result in decisions that are more problematically biased, not less” (Johnson 15). To draw a parallel with her earlier argument, although often not intentionally, judges bring their own assumptions or “canons” to the courtroom. These assumptions can be conscious (like a method of case study or interpretation) and unconscious (like racial bias). Assumptions like these inevitably affect a judge’s ruling, and consequently, the lives of others. 

The AIPIE model makes me curious about the role algorithm programs like COMPAS play in influencing a judge’s assumptions. Specifically, how does their perception of such a tool being objective and value-free reinforce or challenge their own biases? Does disproportionately labeling black defendants as future criminals not only perpetuate oppressive biases and systems structurally, but also reinforce them within the judge’s minds? On the other hand, as much as these algorithms can be flawed, does assuming their objectivity ever meaningfully alter the perceptions of a particularly biased judge or practitioner? In broader terms, what is the nature of the interaction between the inherently biased judge and the inherently biased algorithm? Could their collaboration (rather than independence) result in a more unjust outcome?

Along similar lines, I wonder about the interaction between judges and probabilities (as described on page 17). While not directly related to Johnson’s discussion of probabilities, I wonder how the judges use the risk-assessment score (which is only accurate to a degree of certainty) in their rulings. The emphasis on page 16 that a confidence rating is not a “determinative prediction” seems to suggest that there is a meaningful difference between the two. In practice, is there? Are judges typically mechanical and consistent in their use of the risk-assessment score (which fails to recognize the confidence of error)? If these scores can be used as more of a “guide” than an ultimate decision-maker, would that empower judges to further confirm and defend their own biases?