AI’s Next Scientific Test Is a Chain of Custody for Ideas

A mathematical result is not valuable merely because an answer appears correct. Its value comes from the chain of reasoning, the assumptions beneath it, the ability of others to inspect it, and the confidence that it adds something genuinely new. As AI systems move from solving curated problems toward participating in research, that chain of custody may become as important as the discovery itself. From benchmark scores to research claims In [a report describing ten claimed advances in mathematics and theoretical computer science](https://openai.com/index/ten-advances-in-mathematics/), OpenAI presents its systems as contributors to concrete research outcomes rather than merely competitors on established evaluations. The significance of that claim depends on details that require scrutiny: what was new, how much came from human collaborators, which tools were used, and whether independent experts can reproduce the conclusions. That caveat is not a dismissal. Even a modest number of legitimate machine-assisted results would mark an important shift. Benchmarks usually ask models to recover answers selected or known by humans. Research asks them to operate where the answer may be unk