Comment: When science outruns scientists

AI could dramatically accelerate scientific discovery. Metascience must ensure that speed produces useful new knowledge rather than noise.

In the furore over frontier AI labs developing systems that turn out to be reasonably capable hackers1, it’s easy to overlook some of the more profound advances enabled by AI. Take for example a recent announcement by Google, launching ScientistTwo2, essentially a fully automated PhD researcher, or Anthropic’s note detailing how their AI found a novel enzyme system reminiscent of the 1980s discovery CRISPR, which has become a keystone in gene-editing techniques. Or how two weeks ago OpenAI models solved the Navier-Stokes Millennium Prize Problem3. We’re left to consider that the most consequential application of artificial intelligence may not be writing emails, generating videos, or even hacking websites. It may be that AI can be used to improve the bedrock of human advancement, science itself.

The UK Government’s AI for Science Strategy argues that Britain has a chance to “supercharge our scientific productivity” and establish leadership during an extraordinary period of innovation4. Its first mission is audacious: develop trial-ready drugs within 100 days by 2030. Further missions are promised during 2026.

This ambition no longer sounds entirely fanciful. Cambridge researchers have reported a ten-fold acceleration in identifying potential Parkinson’s treatments5. Google DeepMind’s AlphaFold6 transformed protein-structure prediction, its database expanding access to predicted structures for more than 200 million proteins and, in the process, winning its originators a Nobel prize7. Liverpool’s £81-million Materials Innovation Factory8 has combined robotics, artificial intelligence and chemistry in closed experimental loops; the university now plans the £100-million AI Materials Hub for Innovation9. Google has published research that maps the predicted impact of all 9 billion possible single-letter genetic changes across the human genome using AlphaGenome Atlas10, while developing a new AI weather forecasting model, WeatherNext 311, promising to extend the long-term accuracy of forecasts with impacts in sectors like transport and logistics, agriculture and energy production.

AI is also making significant inroads into the field of mathematics with systems achieving gold-medal-level performance on International Mathematical Olympiad problems12. The rate of discovery has exponentiated from a small handful posited at the start of this decade to dozens of new mathematical discoveries made in 202613. AI tools to support scientific work are proliferating with platforms like Perplexity AI, NotebookLM, Claude Science, GPT-Rosalind and many others enriching the choices available to AI-savvy scientists in almost any field of research.

The fuel for this machinery is not simply advances in AI architectures and algorithms, but data – AlphaFold depended on decades of carefully curated structures in the Protein Data Bank14. Yet laboratories and companies hold vast quantities of “dark data”: failed reactions, negative results, instrument logs and abandoned datasets that were never published or made machine-readable. The UK Government strategy proposes pilots to recover such material so that the negative results, the failures and the mistakes, so often put aside and forgotten by human investigators, can be used to teach an AI where not to look or where to find hidden insights.

But acceleration brings a paradox. AI can increase the production of both knowledge and plausible-looking rubbish. Researchers have already identified an explosion of formulaic biomedical papers mining open datasets for fragile correlations15. Another study found 139 suspicious GPT-associated papers indexed by Google Scholar16. Fields with large, accessible datasets are especially easy to industrialise: the same openness that enables valuable discovery can enable mass-produced inference.

This is why Britain needs metascience, the application of scientific methods to science itself. The UK Metascience Unit, founded jointly by UKRI and the former Department for Science, Innovation and Technology (DSIT), now works across UKRI and its successor department, the Department for Business, Innovation, Science and Trade (BIST)17. Its remit includes experiments in funding allocation, better measures of research quality including responding to AI-authored grant proposals and AI assessments, reproducibility, the changing work of scientists, and expanding the research into the topic of metascience itself.

The central question is no longer whether AI can do science, but what human scientists must become if or when it can.

If advances in AI continue, the scientist’s role will shift from performing every laborious step to that of formulating hypotheses, choosing the consequential questions, designing experimental constraints, judging the evidence and accepting responsibility for the results if trust in science is to be maintained. This will change how we train scientists of the future, emphasising causal reasoning, experimental design, understanding statistics, research integrity and the ability to review and challenge machine output. A doctoral student can already generate an initial literature review in an afternoon, a task that might have taken a year in the pre-AI world.

Validation must also scale to meet the increasing volume of discoveries. AI-generated science should carry provenance: the data, code, model version, prompts, uncertainty and experimental record behind it. High-stakes findings should enter automated replication queues, benchmarked against independent laboratories and adversarial testing before human attention or public money is committed. Journals and funders may need to assess executable evidence packages rather than polished manuscripts.

If advances continue at their current pace, it’s likely that discoveries will become too abundant for human experts to assimilate, so science will require triage. Management of this decision-making process is incredibly sensitive – the idea that some crucial advance might be missed because a triage filter considered it unimportant is worrying, as is the converse idea, that AI introduces misinformation and misdirection into the scientific process. But designing filters that recognise the potential importance of a discovery and highlight independent reproducibility, uncertainty reduction and social value in its assessment is no easy feat.

Managing access to scientific advances raises another issue – how can scientific discovery remain equitable across [nations]? How do we avoid creating AI-powered “science superpowers” who hoard important advances to national benefit? To some extent this has always been a problem but will be amplified beyond recognition by AI. Control of AI is already a political issue with recent events demonstrating that access to frontier AI platforms can be withdrawn at almost zero notice, prompting many governments to consider how they might retain sovereign control of AI18.

The danger is not that machines will make scientists obsolete. It is that we automate discovery while leaving validation, incentives and institutions stuck in the twentieth century, introducing contentious controls on how we find, and who receives, the new discoveries. The study of metascience offers a lifeline, encouraging humans to think about the scientific process itself and how it should be adapted. If AI is to accelerate science, science must first learn to steer itself.