Jessica Hullman, a computer science professor at Northwestern, has written an essay that reframes the AI-for-science debate away from familiar territory — will AI replace scientists? — toward a more uncomfortable question: did we already hollow out our concept of knowledge before AI arrived? The essay's central argument is that the instinct to let AI drive science isn't a forward-thinking embrace of new tools but the logical endpoint of what philosophers call the "value-free ideal" — a post-WWII framework holding that only epistemic criteria, not ethical or social values, should determine what evidence shows. Under this ideal, a correct proof is a correct proof regardless of who or what produced it, and understanding becomes optional overhead rather than the mechanism through which knowledge is actually made. Hullman uses several live examples to ground the argument. Terry Tao's thought experiment — would you take a cancer treatment produced by AI with a formal certificate of correctness and a passed phase 3 trial but no human understanding of how it works — is reframed not as a question about patient desperation but about what happens when the AI "passed the trial by exploiting a weakness in the trial itself." Tom Dietterich's concern about arXiv papers submitted by authors who cannot explain their own claims is repositioned as an accountability problem, not a gatekeeping one: when claims fail, there is no one to consult. The historical sweep is the essay's load-bearing structure. Hullman traces the value-free ideal from late 19th-century scientists worried that public celebration of electric lights would crowd out respect for basic research, through the post-WWII social contract for science (public funding and autonomy in exchange for societal benefit), to the present moment where AI agents running the scientific method at scale look like "the long-awaited fulfillment of nineteenth century scientific ethics." The math case study is particularly sharp. Hullman concedes that math is where correctness can be most cleanly separated from understanding — a proof verified by AI is still correct. But she argues that even here, values are load-bearing: Fermat's Last Theorem became central not because of inherent importance but because Fermat claimed a proof he never provided, enticing others, and it was considered solved after peer review caught an error in the original proof — decades before formal verification arrived. The space of provable theorems is immense; "epistemic taste" is what makes it navigable. The essay's most provocative claim is that the output-first camp isn't making an AI argument at all — they're making a philosophy-of-science argument that was already losing among philosophers before GPT existed. The mainstream in philosophy of science has rejected the value-free ideal, recognizing that values necessarily enter when we weigh consequences of accepting or rejecting hypotheses. The AI debate just gave the ideal's defenders new ammunition. What Hullman is ultimately warning about is institutional drift: if AI takes over problem selection and solution generation, we stop producing people who can judge what's interesting, what's aligned with human needs, and what to do when results fail. The risk isn't that AI gives wrong answers — it's that we lose the capacity to notice when it does.
