Maths dispute puts scientific trust in AI under scrutiny

OpenAI’s claim that its agents solved a mathematical problem relating to the Navier-Stokes equations in 88 hours has prompted questions about research confidentiality. Mathematician Tristan Buckmaster questioned whether his team’s unpublished work, stored in OpenAI’s Codex tool, could have contributed. He stressed he was not accusing anyone. OpenAI denied accessing the work before publication and later said his recent prompts could not have influenced its system.

The dispute highlights broader concerns about originality, accuracy and accountability in AI-assisted science. Google DeepMind’s AlphaFold demonstrated the technology’s potential by predicting protein structures. But its subsequent claims about new crystal structures drew criticism from chemists who found familiar variations rather than strikingly novel compounds. DeepMind maintains that it checked each structure for novelty.

AI-generated writing is also affecting publishing and peer review. Stanford researchers found up to 17% of referee-report text at major computer science conferences had been substantially modified by a language model. A Cornell team found around 147,000 pieces of phantom research cited last year, raising concerns that fabricated references could misdirect research and waste public funding.

Publishers are tightening safeguards. Elsevier uses integrity signals to flag submissions for human scrutiny, Nature’s journals prohibit AI-generated images, and arXiv penalises invented references. Researchers and industry leaders emphasise that scientists must understand, validate and take responsibility for AI-assisted work rather than treat automated output as established fact.