In brief
A controversy erupted when OpenAI announced it solved the Navier‑Stokes equations after a team of mathematicians claimed their research might have influenced the company. The dispute highlighted deeper fears that AI could compromise scientific integrity by producing inaccurate or fabricated results. Researchers note that AI’s tendency to hallucinate is already seen in journals, helping create phantom references and exaggerated findings. Leading publishers are now developing safeguards to ensure AI remains a tool rather than a source of deceit. Amid the chaos, experts argue that human oversight and transparent validation are essential for trustworthy AI‑augmented research.
A UK‑based data vault called Basecamp Research stores one hundred billion genes in a cold‑storage room that once housed a watch factory. Dr. Oliver Vince, the venture’s co‑founder, claims that a couple of sample boxes hold ten times more information than all text on the internet. The company intends to feed this DNA data into an AI that could read biological language, mirroring what an American and a Chinese tech giant do with the internet’s text.
While some academics see AI as a boost to science, others worry the same technology that unveiled new roles for proteins may also blur the line between fact and fiction. The latest flashpoint involves OpenAI’s claim that it solved a long‑standing mathematical problem, the Navier‑Stokes equations, in a little over a day. The institution behind the challenge said it had the “very plausibly” completed the problem, yet soon after, two New York University math professors declared that their own work—stored in OpenAI’s coding tool—could have influenced the company’s result. The accusations were not that a hack had happened, but that the invisible feed of data between AI systems could undermine confidence in the process of verification.
OpenAI answered that no developer or AI had accessed either team’s drafts before the announcement. It later insisted that any data the two months before were incapable of shaping the model. Still, the episode sheds light on a broader risk: the possibility that AI can produce or exaggerate scientific findings, thereby muddying the legitimacy of peer‑reviewed work. Research on a hallucinatory tendency in large language models shows that a recent survey found that around ninety‑percent of biomedical papers scraped for AI fingerprints already carry the language signatures of the models that write them. If unchecked, this could flood journals with impossible references and equations that never exist.
Other high‑profile cases include Google DeepMind’s success at predicting protein structures with AlphaFold, a tool that accelerated discoveries in biochemistry. Yet a group of UC‑Berkeley chemists examined the newly generated crystal structures and found that only a fraction were truly novel, with many resembling known arrangements. The firm’s defense is that novelty depends on subtle changes sometimes unrecognised by human eye, while critics argue that generative models risk prioritising what they think are breakthroughs without robust experimental follow‑up.
In response to these challenges, major publishers and pre‑print sites are setting up new protocols to detect AI involvement. Elsevier has identified a hundred signalling patterns that trigger human review. Some research outlets reject documents that contain AI‑generated images or automatically generated author lists. Others punish authors who submit fabricated bibliographies for a year, underscoring the need for tighter controls as AI becomes increasingly intertwined with science.
Some AI proponents reassure that the technology merely extends human capability. They argue that responsible use—where researchers remain in charge—can still allow legitimate advances. One notable story is that of a biologist who, aided by machine learning, uncovered the surface protein that allows sperm to bind to zebrafish eggs. The research process still required many manual tests, proving that AI can augment rather than replace human insight. Yet the underlying doubt remains: can we rely on systems that routinely hallucinate? Is there still space for a responsible inoculation of AI into the research pipeline, or is the scientific method at risk of being rewritten by black‑box algorithms?
