OpenAI announced on September 8 that it had solved a Millennium Prize Problem, the Navier-Stokes existence and smoothness problem, using 10,000 AI agents working for 88 hours. Within hours, NYU mathematician Tristan Buckmaster accused the company of following a suspiciously similar path to his own unpublished work, stored in OpenAI's Codex tool. The accusation: OpenAI may have accessed his drafts, accidentally or intentionally, to inform its own solution. OpenAI's response made things worse. The company admitted it "cannot rule out that de-identified data derived from product usage helped improve our models." That is a carefully lawyered sentence. It means: even if your raw prompts are not fed directly into training, the patterns and reasoning your usage generates may be.
For European producers, writers and anyone feeding unpublished material through an AI tool, the implications are plain. A treatment you ran through ChatGPT for notes. A budget you checked with a spreadsheet plugin. A casting brainstorm you asked the model to sharpen. All of it may be leaving traces that surface elsewhere, "de-identified" but not gone. OpenAI's Terms of Service promise that "Customer Content" is not used for training if you opt out. What they do not promise, and what this scandal makes explicit, is that the hidden reasoning tokens generated when the model processes your prompt, and the telemetry your session produces, are not covered by the same rule. The line between your content and the system's internal state is not where you thought it was.
Twenty-five Fields Medal winners signed a letter calling AI "misaligned" with the values of open intellectual work. The mathematics community is furious. But the lesson applies wherever unpublished work meets commercial AI. If a mathematician's years of notes can be absorbed into a model without his knowledge, so can a screenplay. So can a pitch deck. So can a production schedule. European law is better here than American law, but enforcement is slow and the remedies are after the fact. The practical move is caution. Assume anything you give a US AI provider may influence the model you share with everyone else.