OpenAI says its AI agents solved a longstanding mathematical problem involving the Navier–Stokes equations in 88 hours. The organization responsible for awarding a prize for a solution said the problem appeared to have been solved, while noting that the work would need time to assess.
Hours before OpenAI’s announcement, New York University mathematician Tristan Buckmaster raised concerns about whether research by his independent team could have reached the company. Buckmaster and Levent Alpöge, a mathematician who also works for Anthropic, had been investigating the same problem and kept drafts in OpenAI’s Codex programming tool. Buckmaster said Alpöge had heard that news of their progress had “reached” OpenAI, but did not explain how. He questioned whether their work might have been used, while stressing that he was not accusing anyone of wrongdoing.
OpenAI said nobody at the company, either a person or an AI system, had access to the team’s work before it was published. The company initially said it could not rule out that anonymized product-use data had helped improve its models, though it considered that unlikely. Days later, it said Buckmaster’s Codex prompts from the two months before the announcement could not have influenced the system.
The episode has drawn attention to a broader concern: AI may weaken confidence in how scientific findings are produced and assessed. Alongside questions about originality and accuracy, researchers worry that AI-generated errors could become difficult to distinguish from reliable results.
AI has already demonstrated significant scientific potential. Google DeepMind’s AlphaFold predicted the structures of 200 million proteins in a year, a task that would have taken hundreds of millions of years using laboratory experiments, according to the account. DeepMind made the results freely available to researchers, and the project’s director and lead scientist received the Nobel Prize in Chemistry.
But claims about AI-generated discoveries have also prompted criticism. DeepMind said its model identified more than two million new crystal structures, with around 380,000 considered promising enough for laboratory production. Chemists Anthony Cheetham and Ram Seshadri of the University of California examined a random sample and said they had found no genuinely innovative compounds, describing many as minor adaptations of known materials. They did not say the method was wrong, but argued that the findings required more rigorous review and validation. DeepMind rejected the criticism, saying it had checked the compounds’ originality and that small changes to known materials can still produce useful substances.
A further concern is the volume of low-quality papers and fabricated references that AI tools may generate. In an analysis of more than one million biomedical papers, computer scientist Dmitry Kobak of Ghent University found signs of AI use in nearly 90% of articles by December of the year covered. AI-detection tools are imperfect, however, and the finding does not establish that every flagged paper was written using AI. Kobak warned that biases in language models could spread through scientific literature.
Peer review may also be affected. A Stanford-led analysis of reviewer reports submitted to four major computer science conferences found that up to 17% had been substantially altered by language models. Separately, researchers at Cornell found that AI-generated bibliographies included nonexistent studies; about 147,000 such works were cited last year, according to the report.
False or unreliable research can lead scientists to pursue dead ends, waste public funds and delay the benefits of legitimate work, including medicines and sustainable technologies. Publishers are responding with additional checks. Elsevier’s scientific-integrity team has developed around 100 indicators to help identify papers that need closer human review. Nature journals prohibit AI-generated images and advise reviewers not to enter manuscripts into AI systems. The preprint platform arXiv suspends researchers for a year if they submit fabricated references.
Researchers also warn that AI systems could overlook valuable findings by treating unexpected results as anomalies. Penicillin, radioactivity and pulsars were discovered through observations that might have been dismissed. DeepMind’s Pushmeet Kohli argues that AI should not be abandoned, but says human oversight is essential: researchers must understand what a system has done and take responsibility for its conclusions.
There are examples of AI supporting research when used under that kind of supervision. Andrea Pauli of the Vienna BioCenter spent nearly a decade studying how sperm enters an egg. Her laboratory used AI as part of work that identified a protein on the surface of zebrafish eggs, suggesting a key-and-lock-like mechanism for sperm entry.
The dispute over the mathematical result remains unresolved in the account, but it illustrates the stakes of AI’s growing role in research: alongside its ability to help scientists make progress, the technology raises questions about attribution, verification and trust.
Sources: Folha Ciência.