OpenAI's math claim sparks dispute over data privacy in AI research

OpenAI announced last month that its artificial intelligence agents solved a long-standing problem related to the Navier-Stokes equations in 88 hours.

The company stated that its system cracked a major unsolved mathematical challenge, a claim that prompted the organization administering the associated prize to say the problem had "apparently been settled." That body noted, however, that a full review of the work would take time.

Hours before the official announcement, Tristan Buckmaster, a mathematician at New York University, raised concerns about the process. He had been collaborating with Levent Alpöge, a mathematician employed by Anthropic, to work toward a similar goal. Their drafts were stored in OpenAI's coding tool, Codex. Buckmaster said his co-researcher was tipped off that news of their progress had "reached" OpenAI, though he did not specify the method of transmission.

Buckmaster asked whether his research, kept in Codex, could have fed into OpenAI's own work, while stressing he was "not accusing anyone of anything." OpenAI responded that no human or AI at the company saw the other teams' work before publication. Initially, the company said it could not "rule out" that anonymised data from product use might have improved its models. Days later, it clarified that Buckmaster's Codex prompts from the two months prior "could not have influenced the system in any way."

The dispute illustrates a wider concern that AI tools may be undermining the integrity of the scientific process. While AI models like Google DeepMind's AlphaFold predicted 200 million protein folds in one year, saving centuries of experimental work, other applications have drawn criticism. DeepMind claimed to have identified more than two million new crystal structures, releasing what it described as "nearly 800 years worth of knowledge." However, chemists Anthony Cheetham and Ram Seshadri at the University of California examined a random sample and concluded they had "yet to find any strikingly novel compounds," arguing many were trivial adaptations of known materials.

Beyond specific discoveries, the volume of AI-generated text is straining academic institutions. In 2024, Prof Dmitry Kobak of Ghent University searched over one million biomedical papers and found that nearly 90% showed AI fingerprints by December of that year. He warned that biases in large language models could permeate the scientific literature. A separate study by Weixin Liang and James Zou at Stanford University found that up to 17% of referee reports for major computer science conferences had been substantially modified by language models.

Citation practices are also coming under scrutiny. Researchers at Cornell University found that AI tools were inventing bogus references, with approximately 147,000 pieces of phantom research cited last year. These errors risk sending scientists down blind alleys and wasting public funding.

Publishers are responding with new controls. Elsevier has developed 100 signals to flag papers for closer human scrutiny, while Nature's journals forbid AI-generated images and instruct referees not to upload manuscripts to AI systems. The website arXiv bans researchers for a year if they submit invented references. Elsevier vice-president Prof Sarah Main said the goal is to ensure the researcher remains in the "driving seat." Dr Pushmeet Kohli of DeepMind emphasized that agents are imperfect and that humans must take accountability for understanding what the machine has done.