Answers are grounded in this paper and cite their sources.
The theories that built the periodic table and the quantum atom were mostly wrong when first proposed, and wrong in productive ways. Prout, Mendeleev, Balmer, Planck, Bohr and Lewis were mistaken about mechanism, or silent on it. Yet each organized the evidence of the day, made checkable predictions, and left structure behind that a better theory kept. This white paper argues that the most valuable role for artificial intelligence in science is to amplify that human capacity rather than replace it: machines that build, keep and retire provisional "scaffold" theories, under goals, criteria and judgements that stay with people. That role was described by the founders of computing and artificial intelligence between 1945 and 1962. We make three contributions. We identify six human characteristics behind productive error in the history of atomic physics and chemistry, and map current AI capability against them. The map shows strength in law and classification recovery and an undemonstrated capacity for the abductive jump and for keeping a knowingly incomplete theory. We argue that AI discovery is evaluated against the wrong target. Benchmarks score recovery of a true law, never the scaffold value of a false one. We propose six scaffold-value criteria drawn from the historical record, each computed by code. We outline a field-neutral benchmark and a governed scaffold engine. The benchmark has counterfactual worlds with hidden laws, staged evidence, sealed prediction registration, calibration against known-wrong reference theories, and pre-declared success and failure gates. In the engine, the machine proposes rival provisional theories while people set goals, declare the non-negotiable core, choose criteria and ratify what enters the record.
No takes yet. Share an insight, caveat, or question.
Painter et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: