The EU’s AI in science plan gets a mixed reception

This week, the EU kicked off its push to lead the world in artificial intelligence-driven science at a summit in Copenhagen, officially launching a bundle of measures to help scientists create new digital tools to aid their research. 

Launching the Resource for AI Science in Europe (Raise) on November 3, European Commission officials expressed the hope that Europe, having fallen behind the US and China in the creation of large language models, might take a lead in applying AI in research. 

Scientists have used forms of AI such as machine learning to help in their work for decades, but a surge of attention and investment in large language models and systems that exploit them, such as ChatGPT, has raised the stakes. The conversation now is about how AI tools might help scientists formulate hypotheses, automate experiments or summarise literature, for example, although the jury is still out on how transformational they will really be. 

Raise is how the Commission intends to deliver its commitment to create a “CERN for AI,” first promised by Ursula von der Leyen in July 2024 while seeking a second term as Commission president. But unlike CERN, Raise will be a distributed throughout Europe, linking existing infrastructures.

Raise “is the platform that will empower a biologist in Copenhagen, a climate scientist in Warsaw, and, of course, a humanities scholar in Rome, to all have access to world-class, large-scale AI,” Maria Cristina Russo, a director in the Commission’s research and innovation directorate, told the conference. 

The aim, she went on, is to outpace the US and China. “They're strong on AI, but they don't have [. . .] a strategy and dedicated actions on AI science.”

Funding first

The core of Raise is a €107 million series of grant calls in Horizon Europe, the bloc’s €93.5 billion research programme. There should be a €30 million call in the first quarter of 2026 to develop foundation models in areas such as materials science and climate change. Meanwhile, €3 million has been set aside in 2025 to help set up a secretariat to oversee the push. 

Several other calls are planned for Horizon’s 2026-27 work programme, due to be released next year. There will be €28 million for “networks of excellence” in materials science, agriculture and environmental science; €30 million for a doctoral network; and €32 million for laboratory automation. 

Raise won’t create a new physical institute, Russo said. Rather, it’s about “connecting the dots” between existing scientists. There’s also a promise to strengthen European coordination on scientific datasets. 

Aside from these Horizon Europe calls, €600 million will also be diverted from the programme to fund the EU’s gigafactories, huge clusters of AI-focused chips that Brussels hopes will provide the hardware for companies and scientists to train new models. 

Investment lags

However, while attendees at the conference welcomed the planned calls, some are concerned that Raise remains something of a grab-bag of grants, without a clear strategy or enough scale. 

“I listened to three separate sessions on Raise [. . .] and I still have no idea what it is in concrete terms,” Nathan Benaich, one of the authors of the State of AI report, an annual analysis of the industry, told Science|Business.

EU investments in AI are “perpetually 10-100 times too small,” said Benaich, who runs a venture capital firm specialising in AI, and who spoke at the conference. 

During the event, he told the audience that “European policymakers are neither playing to win nor playing not to lose, but they’re playing to not lose their own jobs,” garnering applause from the audience. 

“Raise is still at a stage where it can be anything to anybody, a process rather than a structure,” said Thomas Jørgensen, director for policy coordination and foresight and the European University Association.

To be sure, closer European coordination was needed on AI in science, he said. But what was precisely needed, be it better data, more computing power, or talent, differed between different fields.

“Visionary document”

Others at the conference were more supportive of Raise. 

“It's a quite visionary document of how we have to move ahead with AI and science in Europe,” said Kurt Deketelaere, secretary general of the League of European Research Universities, during a panel discussion. But the money matters. “If we want this to be successful, we will need timely and sustained funding,” he said.

Jan Palmowski, secretary general of the Guild of European Research-Intensive Universities, said he too was supportive of Raise, but echoed the need for a different scale of funding. 

“I understand we're talking at the moment about a pilot, but of course, when it comes to the full implementation, we need much more ambitious funding,” he said during a panel discussion. 

Slow and inflexible?

Some attendees in Copenhagen expressed concern that Horizon Europe is far too sluggish and rigid to fund such a pressing area. 

The bulk of the calls that make up Raise will be released in the Horizon Europe 2026-27 work programme, which means money might not actually be disbursed and projects started until 2028-29. 

“One thing I've been missing is speed,” said Thomas Riisgaard Hansen, managing director of Digital Research Centre Denmark, during a debate session. He said he wished the conference had been like an “Apple keynote,” with Raise money available “tomorrow.” 

“How can we get this sense of urgency?” he asked delegates.

Related articles

• State of AI 2025: five key charts for Europeans
• EU launches AI strategies to boost competitiveness and science
• AI in science: is it useful?

Another problem with using Horizon Europe to fund new AI tools for science was its inflexibility, said Anna Kreshuk, a machine learning expert at the European Molecular Biology Laboratory. 

Horizon Europe projects running now were defined before the launch of ChatGPT, for example, but there was limited flexibility to change course in reaction to events. “You have it all defined,” she said. “You can't say, well, the field has moved so much, we should be doing a completely different thing now.” 

This is a criticism that the Commission has taken on board, with research Commissioner Ekaterina Zaharieva saying she wants to reform the programme to allow scientists to switch tack mid-project in response to evolving science. 

Moonshot

In the longer term, the Commission wants to expand Raise in the next EU research Framework Programme, which starts in 2028.

That programme should also include a “moonshot” project on “next-generation AI.” Although details are scant for now, the idea is that it will look at forms of AI beyond large language models, said Liviu Stirbat, head of unit for AI in science at the Commission’s research and innovation directorate, during the conference. 

“It speaks about the next paradigm, going beyond large language models, going to physics based, or science-based AI,” he said. “That's really the kind of longer-term perspective that we take, not really trying to chase the train now, but preparing to lead for the next iteration.” 

Notes of caution

As well as enthusiasm, the conference also heard several notes of caution about AI in science.

Although it could help in areas such as cancer detection, “we must be very aware of the fact that AI can also produce fabricated information or hallucinations, which may distort science and public trust,” said Mikkel Leihardt, director general of the Danish Agency for Higher Education and Science during one session. 

“AI cannot and should not replace scientific curiosity,” said Danish science minister Christina Egelund in a speech. 

A survey released on November 4 of more than 3,000 researchers globally by the publisher Elsevier found that 58% have used some form of AI tool for work, up from 37% last year. 

Researchers are on the whole positive about AI’s ability to save them time, for example by finding and summarising the latest research. But a slight majority say they would not use it to come up with hypotheses, or to design experiments. Thirty-nine percent said AI was “unreliable,” versus 22% who said it was “trustworthy.”