Against Interpretability for Bio AI

This post was submitted by a SAIRC member either as a recommended read or student-created post. All credit remains with the original author.

A contrarian argument from the Sociotechnical Studio at Ginkgo Bioworks: AI models of biology should not be held to the same interpretability standard as conventional scientific models. Drawing on biology's long experience with messy, context-dependent systems, the piece warns that "too much interpretability can be a trap" — pushing researchers back toward tidy, gene-centric narratives that oversimplify what is actually going on — and suggests judging these models on predictive power and data foundations instead.