Can we trace AI contributions to scientific breakthroughs?

On 8 September 2026, OpenAI announced it had solved one of the Millennium Prize Problems, showing that the 200-year-old Navier–Stokes equations "can break down under certain conditions" and so are not reliable for real-world fluids. OpenAI said the result, which cost "several million US dollars," was validated by an automated verification method it calls an emerging standard for mathematical rigor. Twelve hours earlier, mathematician Tristan Buckmaster of New York University, writing for himself and Levent Alpöge of Anthropic, posted that the pair had reached a partial solution with help from both companies' AI tools, and suggested their exchanges with OpenAI's Codex agent "could have been crucial" to OpenAI's breakthrough. OpenAI denies this. Nature's editorial calls the episode "a wake-up call."

The editorial's reasoning: AI systems are "already well on the way to acquiring and digesting all of digitized human knowledge," but the neural networks underneath are "black boxes" that "do not necessarily keep track of what they learnt, from where or how" — so when a model reaches a breakthrough, tracing the "starting hints" back to their human source can be "almost impossible." Because "information to verify those concerns has not been published," the dispute cannot be settled either way. The editorial's remedies follow from that opacity: switch data sharing from opt-out to opt-in by default, audit AI agents independently to curb "unsanctioned AI agent behaviour," have academic institutions tighten agreements covering what platforms may train on, and bring AI companies into the Leiden declaration's attribution pledge.

This sits beside Can AI-generated proofs ever replace human mathematical understanding?, whose pledge the editorial explicitly wants extended to the AI companies themselves, not just the mathematicians using their tools. It cuts against the self-reported productivity framing in Are AI agents now doing more research work than humans?, treating OpenAI's own verification claim with suspicion rather than taking it at face value. It shares with What stops AI from discovering science without human help? a concern that AI systems' internal opacity, not just their scale, is what makes them hard to govern.

The excerpt does not establish that Buckmaster and Alpöge's interactions actually fed OpenAI's model, or that OpenAI's verification method is sound — both remain disputed and unpublished. What it does establish is that no mechanism currently exists to trace or credit the human contributions behind an AI-generated mathematical result, and that this is a Nature editorial's institutional judgment, not yet a documented case of misattribution. The implication, held at that strength, is that attribution norms for AI-assisted discovery need building before the next disputed claim, not after.