Answers are grounded in this paper and cite their sources.
An AI research agent can run the wrong experiment perfectly. This paper examines a class of research failure that occurs after implementation is correct and the hypothesis is well specified: frame failure. The researcher asked a coherent question, built an appropriate measurement, and obtained a legitimate result, but the question itself was not the question that mattered. We distinguish three kinds of failure. An implementation failure means the experiment did not execute as intended. A hypothesis failure means the stated hypothesis was tested and did not survive. A frame failure means the experiment was valid, but the framing that determined what counted as the relevant question was wrong. The third category matters more as AI systems become capable of proposing, implementing, executing, and evaluating experiments at high speed. Three concepts organize the problem. Legibility bias is the tendency for questions that fit existing conceptual and technical machinery to become easier to formulate and test, without thereby becoming more true. Lamp–Map Collapse is the institutional failure in which the territory illuminated by an increasingly capable research system is mistaken for the whole territory of possible inquiry. One-Free-Flip is a falsification discipline for frame changes: a researcher may reject the old frame when evidence warrants it, but the replacement frame inherits the burden of specifying what would falsify it. The resulting partnership is asymmetric in capability and consequence, not in epistemic entitlement. The AI system is especially valuable as a source of accumulated knowledge, precedent, instrumentation, implementation, and measurement. The researcher carries consequence for the direction of inquiry. The framework is derived from research practice, not presented as an experimentally validated result. The objective is not to escape the center of established knowledge. It is to move the center.
No takes yet. Share an insight, caveat, or question.
Guha Swaminathan Kashyap (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: