The BBC starts a fight around Navier-Stokes to draw up an inventory of risks: submerged reviews, revisions written by models, 147,000 references invented in one year.
The BBC's science correspondent, Pallab Ghosh, questions a fear less spectacular than biological weapons: that AI will erode the trust on which research is based. Starting point is the controversy around OpenAI's demonstration of the Navier-Stokes equations, which mathematician Tristan Buckmaster has disputed by questioning whether his drafts, stored in Codex, could have fed into the company's work — which OpenAI denies.
He also recalls the criticism of two chemists at the University of California, who found no truly new compounds in a sample of the two million crystals announced by DeepMind, which DeepMind disputes. More worrying for the author is the pollution of the literature: up to 17% of the text of review reports at computer science conferences has been largely rewritten by a model, according to a team at Stanford, and a team at Cornell estimates that about 147,000 ghost references were cited last year.
Keeping a human accountable
Publishers are starting to react: Elsevier relies on a hundred red flags, Nature journals prohibit submitting a manuscript to an AI, and arXiv suspends authors who cite invented references for a year. Pushmeet Kohli of DeepMind sums up the rule of thumb: understand what the agent has done and take responsibility for it, rather than leaving them to write and submit an article alone.
Read the full article on the original source
Go to the source →