OpenAI's Navier-Stokes Claim Sparks Debate Over AI Role in Scientific Discovery

OpenAI announced last month that its artificial intelligence agents solved a long-standing mathematical problem related to the Navier-Stokes equations. The company stated the solution was reached in 88 hours. This claim has intensified concerns among scientists about the reliability of AI-driven research and the integrity of the peer review process.

The controversy emerged hours before OpenAI made its public announcement. Tristan Buckmaster, a mathematician at New York University, raised questions about how the solution was derived. Buckmaster worked with Levent Alpoege, a mathematician employed by Anthropic, on a similar problem. They used OpenAI's coding tool, Codex, to store their drafts. Buckmaster noted that he was informed their progress had reached OpenAI, though he did not specify the method of transmission.

OpenAI initially stated that no human or AI within the company viewed the other team's work before publication. The firm acknowledged that anonymized data from product use might have improved its models. A few days later, OpenAI revised its position, stating that Codex prompts from the two months preceding the announcement could not have influenced the system. The precise mechanism remains unresolved, but the incident underscores growing tension over AI's place in academic work.

Critics argue that AI may undermine trust in scientific discovery. Dr. Oliver Vince, co-founder of Basecamp Research in London, believes AI will accelerate biological research by processing vast datasets. His firm stores gene samples in freezers at a former watch factory in central London, aiming to build an AI model that interprets biological data similarly to large language models process text.

Skepticism also surrounds the validation of AI findings. In 2020, Google DeepMind's AlphaFold model predicted the structure of 200 million proteins, a task that would have taken hundreds of millions of years in a laboratory setting. The developers received the Nobel Prize in Chemistry for this work. More recently, DeepMind claimed to identify over two million new crystal structures. However, chemists Anthony Cheetham and Ram Seshadri at the University of California reviewed a sample and concluded that many were trivial adaptations of known materials rather than novel compounds.

Publishers are responding to a surge in AI-generated content in scientific journals. Dalmeet Singh Chawla, a science journalist, noted that publishers' integrity teams need to drastically increase their resources to manage the influx of submissions. Prof. Dmitry Kobak of Ghent University found that nearly 90% of a million biomedical papers contained AI fingerprints by December 2024. Separately, researchers at Cornell University identified approximately 147,000 phantom citations in last year's literature, suggesting AI systems are inventing references that do not exist.

Institutions are implementing stricter rules to address these issues. Elsevier has developed about 100 signals to flag papers for closer human review. Nature's journals prohibit AI-generated images and prevent referees from uploading manuscripts to AI systems. The preprint server arXiv bans researchers for a year if they submit invented references. Prof. Sarah Main, an Elsevier vice-president, described the goal as creating research-grade AI where the human researcher remains in control.

Dr. Pushmeet Kohli, who leads DeepMind's science work, argued that rejecting AI is similar to computing without calculators. He emphasized that agents are currently imperfect and require human accountability. The challenge lies in ensuring that AI tools, such as those processing data from particle accelerators or telescopes, do not dismiss valid anomalies. Scientists must understand what the AI has done and take responsibility for the results before they are submitted as true.