AI speeds up science while trust in results is questioned - critical summary review

Artificial intelligence does not run experiments. In a laboratory it does something narrower and, in practice, very useful: it sorts through possibilities. Millions of chemical compounds, billions of genetic sequences, thousands of satellite images. The machine ranks them and points to the few worth a researcher's time and a test at the bench.

That is why AI keeps turning up in stories about battery materials, antibiotics and archaeological sites, and why it reaches people who will never open a scientific paper. It changes which drugs get tested, which materials get built, and how quickly.

The speed gains are documented. What has not moved at the same pace is the ability to check each individual result. Recently a computational biologist at the University of Copenhagen publicly questioned a discovery that the AI company Anthropic attributed to its own software, and a seven-year research project funded by the US military's research arm reported that AI still cannot reliably tell which published studies will hold up.

What AI has already sped up

The clearest cases are about filtering. Microsoft and the Pacific Northwest National Laboratory, part of the US Department of Energy, used AI and supercomputers to narrow a list of 32 million candidate inorganic materials to 18 promising ones in less than a week, a screening job the researchers said could have taken more than two decades by traditional lab methods. The material they ended up testing is a solid electrolyte that could cut the lithium used in batteries by up to 70%.

At the Massachusetts Institute of Technology, James Collins and his team used generative models to design 36 million possible compounds against two bacteria that resist most available drugs, the one that causes gonorrhoea and the Staphylococcus aureus behind MRSA infections. They synthesised 24 in the lab, seven showed antimicrobial activity, and two killed strains resistant to other antibiotics. Both are still in testing.

In archaeology, a Japanese team working with IBM Research used AI to identify 303 previously unknown geoglyphs in Peru's Nazca desert during six months of fieldwork, almost doubling the known total. The study was published in PNAS, the journal of the US National Academy of Sciences.

None of this is marginal. The number of peer-reviewed papers using AI tools has risen sharply since 2015, and more than half of the scientists surveyed by the journal Nature expected such tools to be "very important" or "essential" to research practice. In 2024 the Nobel committee recognised the field twice: physics for the machine-learning foundations laid by John Hopfield and Geoffrey Hinton, chemistry for work on protein design and for AlphaFold, the model built by Demis Hassabis and John Jumper that predicts protein structures.

When a discovery becomes a dispute over authorship

The question of where the machine got its answer is not new. José R. Penadés and his team at Imperial College London spent a decade trying to prove why some superbugs are immune to antibiotics. He put the central question to co-scientist, a tool made by Google, and in 48 hours it reached the same conclusion. His research had not been published, so the system could not have found it in the public domain. "I wrote an email to Google to say: you have access to my computer, right?", he told the BBC. The company confirmed it did not.

The latest case is less amicable. When Anthropic presented the conclusions of its new biology lab, its scientists said AI agents had found promising new enzymes for biotechnology, which the company calls ARTs, and described the work as the "first steps of a biological discovery" taken by the AI on its own. Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen, said he and colleagues had been studying those same enzymes for four years without publishing, and had spent the past three years using Anthropic's models to write code and draft manuscripts, sharing drafts of a thesis and a manuscript along the way. "My concern is that this information was used to train future versions of the models," he said in an interview. Anthropic said in a statement that it was not aware of any previously published work describing the system it found, that Claude was not trained on user transcripts, and that its molecular biology team has no access to them. The company's researchers did not describe the exact prompts they used.

Something similar happened in October, when OpenAI said GPT-5 had solved an Erdős problem, a classic benchmark of mathematical performance. Researchers pointed out that the model had only surfaced solutions already published, including in old papers and other languages that were hard to find. A prominent mathematician noted he had been using the company's AI for some time to try to reach a solution. Since then, says Kevin Weil, the vice-president who leads OpenAI's new science team, the company has leaned on "epistemic humility": "We are spending a lot of time trying to make sure the model says here is something to consider, rather than here is the answer."

The project that tried to score reliability

Scientists publish more than 10 million studies and other works a year. Seven years ago, hundreds of researchers set out to find a faster way to tell which ones would survive scrutiny, in a project called SCORE, short for Systematizing Confidence in Open Research and Evidence, funded by the US defence research agency. The idea, from its then programme manager Adam Russell, was a kind of credit score for science.

The team replicated 164 social science studies. Only about half produced the same results as the originals. When members reanalysed published data using the authors' own code, part of the findings came out different; when they had to write their own code, they reproduced the exact results less than half the time.

The AI trained on all that is not there yet. Its predictions are not random, but they are far from accurate. "It is picking up some kind of signal, but it would need to be much more precise to be used on its own," said Brian Nosek, executive director of the Center for Open Science and one of the project's leaders. Human experts asked to guess which papers would replicate got it right about three quarters of the time.

Retracted papers, and a science that sounds like itself

There is a second layer to the trust problem. According to MIT Technology Review, popular chatbots including ChatGPT sometimes cite scientific papers that have already been retracted, without telling the user the work was withdrawn from the record. One obstacle is that publishers do not use the same labels for retractions and errata, which makes automatic detection harder. Specialists argue for real-time checks against databases such as Retraction Watch.

A paper in the journal Trends in Cognitive Sciences raises a subtler risk. Because large language models are trained on word statistics, the authors argue, they "tend to reproduce conventional, institutionally validated perspectives and writing styles" and create "an illusion of consensus" that treats those norms as the standard of clarity while silencing alternative worldviews. The Royal Society, Britain's academy of sciences, has separately warned that the black-box nature of many AI tools limits the reproducibility of research based on them.

The counterweight, say researchers who work with these systems, is that a proposed molecule is not a made-up citation. "It has to be tested," said Anima Anandkumar, a professor at the California Institute of Technology, about AI results grounded in physical data. The test is the part that does not get faster.

What to do with this information

The useful signal is not the announcement. It is what comes after it. Was the result published and peer reviewed, or only posted on a company website? Has another lab reproduced it? Did the team release the prompts, the data and the code it used? In the Anthropic case, the absence of those details is precisely what left room for contestation, and in the SCORE project the inability to rerun other people's analyses was one of the main sources of failure.

It also helps to ask how far along a result is. Compounds found by AI against resistant bacteria have been tested in lab dishes and in mice, not in people; the researcher behind one of them hopes the first AI-assisted antibiotics could be prescribed by 2030. A screening that takes days replaces a screening that took years. It does not replace clinical trials, replication or peer review, and none of those has been compressed in the same way.

The next time an AI discovery crosses your feed, how much of the story goes past the announcement: a reviewed paper, shared data, or another lab that got the same result?