AI in Science is a marathon. Can Europe lead on trust?
Reflections from the #AI2S26 AI in Science Summit in Dublin on scientific rigour, collective ambition and making AI work for research.
Dublin was a lovely setting for two days of conversations about the future of AI in science. With an excellent venue at University College Dublin and thoughtful organisation, the AI2S26 Summit gave us space to meet, exchange views and look beyond our own projects. There was plenty of interest, a strong mix of public organisations and industry, and a sense that many of the people who could help move things forward were in the room.
Much of the conversation was at a high level, as you would expect. I could see the enablers of AI in science well represented: policymakers, infrastructure providers, technology companies and organisations supporting research. I was less sure that we were hearing enough from researchers themselves. Perhaps this was not the setting for detailed discussions of everyday research. Still, the choices being discussed will shape their work. We need their experience of where AI helps, where it falls short, and what it takes to check its results. They should have a direct role in shaping the priorities and services being developed for them.
Can Europe lead on scientific rigour and trustworthiness in AI? We may not have been first out of the blocks, but we have the knowledge and experience to help get this right. A colleague from the European Commission told me last year that AI is a marathon, not a sprint. That stayed with me. And marathons are rather a European invention, after all.
The question is where Europe should best position ourselves. Can we build the scientific "harness" that allows researchers to examine how results were produced, test their reliability and understand their limitations? And in the end get the accreditation they thrive for? Where is the balance between safeguarding research and giving researchers the freedom to explore? We have the brains and the capacity. Do we have the collective commitment to work across our fragmentation and build the public-private partnerships we need, in ways that preserve trust in science?
This is the role envisioned for RAISE: bringing European capabilities together around the needs of scientific communities. Its success will depend on turning that ambition into sustained support that researchers can access, use and help shape.
Public-private partnerships: reasons to be optimistic
It was excellent to hear the announcement of the CERN–Mistral partnership to develop a single pretrained Large Physics Model – the first foundation model for collider physics. If they get it right, many more scientific disciplines could follow. Of course there were the usual "let's wait and see first" views in the corridors, but I am quite optimistic that this can work. Scientific organisations and industry have much to offer each other. We need to put the effort into building relationships where scientific priorities, openness and public value are part of the agreement from the beginning.
This reminded me of Kaylin Bugbee's presentation from NASA at the "AI with and for Open Science" panel I moderated at Open Science Fair in Madrid in 2023. She described NASA's collaboration with IBM on language models for science, with open science and carefully curated data at the centre. I also wrote about it in my reflections from that panel. The collaboration produced INDUS, an openly available suite of scientific language models released in 2024. NASA reported promising initial search results in its Science Discovery Engine prototype and integration into its data services. It is good to see something we discussed then becoming tools that others can use and build on.
My takeaway: public-private collaboration can deliver. The terms of that collaboration will determine how much lasting value it creates for science.
Publishers and data: have we learned enough from the past?
Against that background, I have to admit that I was unpleasantly surprised by the Commission's industry outreach presented in a few panels. Two of four industry liaison places went to Elsevier and Frontiers. My first reaction was: the familiar players again, with scientific data now at the centre. Are we thinking broadly enough about who should help shape this next phase?
I am very much in favour of working with industry, including publishers. For years, I have argued for publishing models built around valuable services, where content remains open and institutions can change providers. There is room for sustainable businesses that support science without controlling access to its knowledge.
But I heard too little discussion of those alternatives. What are research organisations contributing to these partnerships, and what do they retain? Who can reuse the data (and the model), under which conditions, and what happens if the relationship changes? We should be discussing these questions openly.
Perhaps I missed parts of the conversation. But I left concerned that we could reproduce dependencies that took decades to address in scholarly publishing. With AI, those dependencies could reach much further into the research process.
My takeaway: the enthusiasm for partnerships needs to come with an equally serious conversation about business models, model and data rights and freedom of choice.
One Health: turning a shared concern into shared work
I followed the One Health strand, and it was good to see a global community bringing together different disciplines and perspectives. Human, animal and environmental health are connected, and there was a clear willingness to work across those boundaries.
But I found myself asking where the practical work begins. From what I heard, there was much more discussion of the ambition than of concrete implementation. Is the topic so broad that it becomes difficult to agree on a starting point? Could a few well-defined challenges bring the different communities together around shared data, concrete tasks and results they can assess together? That might also help clarify where AI can make a useful contribution and what is needed to put it to work.
My takeaway: give the collaboration somewhere concrete to start. Agree on problems, responsibilities and milestones, and resource the work needed to connect disciplines and data.
OpenAIRE's contribution: working through the questions together
OpenAIRE was actively involved on both days, through the policy challenge and our workshop on Open Science and Trustworthy AI. A very nice surprise was the diversity of people around the tables, and the presence of so many young people. Different backgrounds and experiences brought different questions, and a willingness to challenge assumptions. It was a real pleasure to work through these issues together.
Day 1: what should policy make possible?
On the first morning, I moderated a table at the policy challenge organised by SCIANCE colleague Eugenia Vilarchao from ESF bringing it all together. We had a lively and knowledgeable group.
Several points stayed with me:
- Set a clear ambition for AI in science. Better, more trustworthy research and public benefit should guide investment. We need to be clear about what we want AI to improve and how we will judge whether it does.
- Support the whole research environment. Data, tools, workflows and human decisions all shape the results. Institutions need resources to provide infrastructure, guidance and training, with access that does not depend on an individual researcher's budget.
- Make responsible use practical. Requirements should reflect the task. Exploring an idea and validating a finding call for different levels of scrutiny. Researchers should feel able to disclose AI use openly.
- Use procurement to shape the market. Collective negotiation can give research organisations more influence over price, data use and access. These decisions will shape our future options.
- Give European coordination continuity. Our table called for a stable mandate and funding for RAISE, connected to existing European networks and national initiatives. We cannot keep rebuilding capacity through disconnected projects.
Day 2: what would make trustworthy AI work?
Our Open Science and Trustworthy AI Working Group held a workshop through SCIANCE, in support of RAISE. Four tables explored transparency and reproducibility, accountability and integrity, legality and rights, and sovereignty and control.
For me, the discussions pointed towards several strategic priorities:
- Build trust into research infrastructure. Documentation, provenance and records of AI-supported work need to be supported by the systems researchers use. Adding another reporting obligation will achieve little unless we also make it feasible.
- Give scientific communities a role in setting standards. What counts as sufficient evidence, suitable data or a reproducible result varies across disciplines. Research infrastructures and communities need resources to develop workable approaches with model providers.
- Make accountability clear across the whole chain. Researchers, institutions, infrastructure operators and providers each have responsibilities. Agreements need to explain who checks systems, investigates failures and corrects the record. Shared responsibility must not become an excuse for nobody taking responsibility.
- Treat evaluation as continuing work. Models change, their uses change, and weaknesses emerge in practice. Testing before deployment needs to be followed by monitoring and ways to report problems. Human oversight also needs expertise and time to be meaningful.
- Protect confidence in research assessment and peer review. Confidentiality, bias and the ability to question decisions matter. AI may help with growing workloads, but we also need to examine the incentives driving submission volumes and the pressure on reviewers.
- Make sovereignty a usable option. European services need to be competitive in performance, affordability and usability. Institutions need the capacity to manage dependencies and retain choices. This should support international scientific collaboration and the global research commons.
- Align incentives with responsible practice. We should recognise the work of validating tools, curating data and documenting methods, alongside new scientific findings. Training will have limited effect if the research system rewards speed and volume while overlooking this work.
These exchanges gave us a substantial foundation for our report and contribution to the RAISE Strategic Research and Innovation Agenda. They also made clear that implementation needs named responsibilities, sustained resources and involvement from the communities expected to put it into practice.
A big thank you to Fatemeh Raji, César Dro, Fotis Psomopoulos, Dorottya Sallai and Manolis Terrovitis for moderating, and to everyone who contributed. Special thanks to Theodora Kavvadia from OpenAIRE for organising the second-day workshop, to Alane Brunschweiger for supporting us.
And we launched the RAISE Observatory
The RAISE Observatory, powered by the OpenAIRE Graph, was also launched in Dublin, with OPIX (our member Athena Research Center & Innovation Centre spinoff) collaborating on the ontology and inference. It is now out for feedback, and we are already working towards the second release.
This is another part of the work ahead: building evidence about how AI is developing across research fields, while being clear about what we can measure and where our understanding is still incomplete. Feedback from research communities will be essential.
I left Dublin optimistic, but impatient to see the next steps. The expertise is there, and so is the interest. What I would like to see now is that collective commitment translated into partnerships, shared investment and practical support for researchers. A marathon takes sustained effort. We need to commit to it together.
When you subscribe to the blog, we will send you an e-mail when there are new updates on the site so you wouldn't miss them.