In brief
Biologists and tech pioneers clash over whether large language models can safely shepherd scientific discovery. Basecamp’s Drake plan to mine DNA for a biology‑talking AI clashes with concerns that AI workers could unintentionally leak research data. OpenAI’s claim of solving Navier–Stokes raised alarm when a mathematician alleged that his coworking prompts may have influenced the breakthrough. Experts warn that AI‑generated papers and citations are polluting journals, jeopardising peer review, and risking public funding being misdirected on dead‑end projects. While some see AI as a tool that can speed breakthroughs, others argue that unchecked algorithms threaten the very bedrock of trust in science.
Scientists at a London‑based data‑storage facility claim that a small number of vacuum‑sealed tubes could outstrip the entire internet’s text in information density. The company, called Basecamp Research, intends to use that genetic goldmines to feed a new AI that can “talk biology” in the same LLM‑style as GPT‑like systems. Yet activism in academia warns that the same AI tools are quietly eroding the scientific process.
The Bump‑up
When OpenAI announced it had cracked a decades‑old maths problem in just a few days, a New York University mathematician sounded the alarm. Tristan Buckmaster, who had been collaborating with a scholar at competitor Anthropic, alleged that their drafts were accessed via OpenAI’s own coding platform, Codex. Though no accusation of data theft is lodged, the claim raises doubts about whether AI‑powered research sits on the shoulders of its peers or inadvertently steals them.
Muddling the Journal Funnel
Researchers across physics, biology and chemical engineering report that AI can fabricate citations, pull in fabricated references, and occasionally publish polyethylene papers that contain no real data. An analysis by a team in New Hampshire found that 90 % of biomedical papers reviewed last year bore an AI fingerprint. Journals blame a lack of human gatekeepers for this “research‑grade slop.”
DeepMind’s Double‑Edged Sword
DeepMind’s AlphaFold, which predicted the 3D structure of millions of proteins overnight, remains a celebrated triumph. Yet the firm’s later foray into crystal‑structure prediction produced two million “new” designs that were criticised for lacking novelty. Critics argue that an AI finding a single altered atom can’t be equated with human ingenuity, and that unchecked novelty claims may poison the literature.
Guardrails for the New Age
Elsevier has rolled out a suite of 100 signals that flag suspicious submissions, while Nature prohibits AI‑generated imagery. ArXiv points an author to a one‑year ban if a fake reference is detected. Still, professionals like Sarah Main at Elsevier stress that AI must stay “in the driver’s seat” and that researchers must remain accountable for what the models produce.
Could AI Detect the Missing Jackpot?
The reliability of AI is tested as it sifts through the deluge of data from particle colliders and telescopes. If an AI system arbitrarily dismisses a serendipitous anomaly—as human scientists historically risked—measurable discoveries may be lost. To counter this, scholars argue for hybrid approaches where AI outline findings, but humans verify the significance and context.
The debate continues: can AI accelerate cures, or will it quietly dissipate the trust that underpins progress? Basecamp’s founder, who survived leukemia, believes a top‑notch AI path could help battle cancer. The question remains: will the benefits outweigh the erosion of science’s foundational bricks?
