A new contribution to the debate over artificial intelligence in science argues that correctness and verification are not sufficient substitutes for human understanding, criticism and accountability. Jessica Hullman, a Northwestern University professor of computer science, says the current argument is too often framed as a choice between obtaining useful answers and preserving an optional human mental state.
Hullman’s essay examines proposals that AI-generated proofs, treatments or papers should be accepted when their outputs can be verified, even if no person understands how they were produced. She argues that this framing becomes weakest precisely when a result fails. A treatment may pass a trial by exploiting a weakness in its design, while a paper whose named authors cannot explain its reasoning leaves no responsible expert available when a claim is challenged.
The issue is therefore not simply whether humans feel intellectually satisfied. Scientific knowledge depends on people and institutions being able to state what a claim rests on, test its boundaries, respond to criticism and revise it when evidence changes. Verification can establish that a result passes a defined check, but it cannot guarantee that the check captures every relevant scientific, ethical or social concern.
The essay connects the debate to the postwar ideal that science should be insulated from contextual values and judged through supposedly impersonal methods. That model helped support a social contract in which researchers received public funding and autonomy in return for practical benefits. Hullman argues, however, that treating scientific judgment as value-free can hide the choices embedded in research: which questions receive attention, what counts as adequate evidence and which risks are considered acceptable.
AI makes those longstanding tensions more visible because systems can produce large volumes of plausible or formally checkable material without assuming responsibility for it. If journals and institutions treat successful verification as the whole of knowledge production, criticism and expert interpretation can be mistaken for inefficient overhead. The result may be a scientific record containing answers that are difficult to integrate, interrogate or repair.
Hullman does not argue that AI should be excluded from research. Her case is that scientific value is created through a social process as well as through correct outputs. Tools can help researchers generate hypotheses, proofs and analyses, but the surrounding system still needs accountable people who understand the assumptions and can answer objections.
That distinction matters as laboratories, publishers and funders decide how to evaluate AI-assisted work. A machine-generated result may be useful and even correct, yet still fall short of knowledge if nobody can explain why it should be trusted outside the narrow conditions of its verification.
