At the Metascience 2025 conference in London, discussions about Artificial Intelligence in science produced cautious views rather than optimistic predictions of wholesale change. Delegates acknowledged recent breakthroughs, including large language models and tools such as AlphaFold, and suggested possible uses like generating hypotheses, replicating computational work, or summarising existing literature. Matt Clancy of Open Philanthropy said new tools open fields but do not amount to a fundamental transformation of what it means to be a scientist. The eu research commissioner Ekaterina Zaharieva described the transformation as impressive and promised an Artificial Intelligence in science strategy soon, following the commission´s creation of a dedicated unit and publication of 15 case studies arguing the technology speeds discovery in life sciences.
Speakers emphasised that many celebrated advances rest on decades of prior research and data. Iulia Georgescu traced the history of Artificial Intelligence in science back to a 1956 theorem proving tool and noted machine learning played roles in physics in the 1990s. AlphaFold was highlighted as arguably the biggest recent gain, with DeepMind winning a share of the chemistry Nobel Prize and Anna Koivuniemi noting the system expanded known protein structures from roughly 200,000 to 200 million and has been used by more than three million researchers. Koivuniemi stressed that AlphaFold depended on high-quality training data produced by structural biologists over more than 50 years, underscoring limits to what current models can do without good data.
Concerns about accountability, quality and unintended harms featured strongly. Researchers remain wary of delegating work they must ultimately vouch for, and tests of literature-review tools by the Columbia Journalism Review found results underwhelming and sometimes alarming. Cost and access are barriers in some countries: Moumita Koley reported limited uptake of large language models in India, where researchers mainly use them to polish writing and worry about future expenses and restrictive journal policies. Other contributors warned of a possible flood of low-value or fraudulent papers generated with models, which could overwhelm researchers and obscure genuine findings. Liudmila Zavolokina reported finding several AI-generated papers falsely attributed to her on Google Scholar, a concrete example of the risks debated at the conference.
