CITP Seminar: Mel Andrews - False Promises & False Premises of Fair Machine Learning

Far and away the most prominent and practically influential approach to AI ethics to date is the paradigm of algorithmic fairness. Machine learning applied in high stakes contexts is an epistemic activity, aimed at inference over real world data and hence real world systems. The paradigm of ML fairness is a second-order epistemic activity, aimed at evaluating the success of a first-order modeling activity; call this a metamodeling technique. Insofar as the fairness framework can offer normative guidance, it is only in virtue of successfully serving its role as metamodeling approach. A case will be made for goal-specific adequacy conditions on such metamodeling exercises; the standard usage of fair ML methods is revealed to be ill-posed in light of these. Applied ethical frameworks like algorithmic fairness ought to be held to the candle of epistemic adequacy.

Bio:

Mel Andrews researches both the promises (and pitfalls) of incorporating AI into scientific pipelines and the scientific nature of AI deployed in socially-sensitive arenas. Current projects evaluate prospective uses of machine learning in peer review, grant review, and metascientific applications, looking to offer guidance to institutions for science oversight in their development of AI use policies. Andrews earned a Ph.D. in philosophy of science from the University of Cincinnati in 2025. Andrews’ work runs across disciplinary divides, drawing on the scholarly traditions of history and philosophy of science, science and technology studies, and formal methods, alongside firsthand knowledge of practices in both laboratory and computer science. They are jointly affiliated with the Princeton AI Lab’s Natural and Artificial Minds initiative and supported by a Sloan Foundation Metascience and AI grant.

Sponsorship of an event does not constitute institutional endorsement of external speakers or views presented.