Voices in EU AI Policy: an Interview with Sebastian Lobentanzer on Europe's Frontier AI Strategy

We asked Sebastian what these two engagements mean for European AI policy, for AI-for-science coordination, and for Helmholtz AI.

Let's start with the Expert Forum on Frontier AI. What is it, and how did you get involved?

Describing it in my own words, it is a mechanism the Commission chose to get rapid input on the current state and challenges of frontier AI in Europe as part of the Frontier AI Initiative. There was a call for expressions of interest in February, which sounded interesting, so I put one in. The meeting was in April; a full afternoon of short pitches and breakout discussions among 100+ participants. I was one of around 20 participants from “think tanks, academia, and other research organisations” alongside at least three other Helmholtz colleagues (but naming them is prohibited under the Chatham House Rule). The Commission organisers set up the programme and took detailed notes to paint a picture of the state and challenges of frontier AI in Europe, which was published in July. The report can be found here.

In terms of policy at the scale of a continent, this is fast. I appreciate the desire by the Commission to found their decisions going forward on a transparent and expert-driven assessment. The opinions I heard in the forum were very candid and substantive. 

However, it also shows that fast in policy does not mean fast in the real world. The six months it took to assemble and summarise this forum are almost an eternity in the AI landscape. The forum surfaced an interesting meta-problem: how should regulation keep up with an ever-changing high-pressure landscape dominated by the richest companies on the planet?

The Forum's report argues Europe has a narrow window of one to two years to secure its position in frontier AI. Do you think that sense of urgency is warranted, and if so, what needs to change first?

I would say that – rather than arguing – the forum’s report is a neutral summary of the diverse perspectives of that meeting. The composition of the meeting alone shows the multi-dimensional nature of the problem, as there were members of industry and academia, economists and biologists, deep learning experts and lawyers, all sharing their perspectives on a global phenomenon. To me, it emerged that we cannot allow ourselves to be paralysed by the complexity of the problem. The pull towards paralysis is strongest for those in governance positions, where it is most tempting to wait for more certainty.

From a scientific perspective, I agree with the report’s assessment that models – in some domains – have moved from struggling with basic tasks to approaching the limits of what our tests can measure (§2). However, translating this naively to highly complex unsolved domains, such as biology, is a fallacy.

Nevertheless, I share the position of urgency; not because I think that any single year is going to be decisive, but because it seems more important to do something than to be caught in endless back-and-forth of (relatively) meaningless details. Europe unfortunately is quite good at letting internal struggles incapacitate decisive action. The approach the Commission has chosen in this instance is the right one: ask the experts, summarise the findings, use them to make good decisions.

What I think needs to fundamentally change to keep pace with the current developments without sacrificing our political values is the isolation of professions. Policymakers, regulators, industry, academics – these are complementary and they need to establish trust and open communication in order to move forward. The natural next step is now to digest and prioritise the forum’s insights into a high-level agenda of action items for Europe’s AI strategy. Above all, the Commission needs to show it is willing to listen to experts and act on their advice in a timely manner.

Your own research is on governing agentic AI systems in biomedicine. Where does that fit into a conversation about frontier AI at the EU level?

There was quite a bit of conversation about domain-specialist systems (for instance in biomedicine) at this meeting. In this context, frontier AI is defined as any AI system that advances the state of the art in some domain. The public is currently equating frontier AI (and, let’s be honest, also just “AI”) with large language models (LLMs). This is only natural, as LLMs have come to great prominence in the last couple of years and are the first mass consumer AI applications to speak of. But it makes you forget that the field of artificial intelligence has a rich history and that many extremely valuable applications are already found in daily life, from finding a route with a digital map to the weather forecast, in which AI models now match compute-heavy physics simulations at a fraction of the cost. At some point, they might even keep the trains on time. There are highly important and exciting frontiers in all domain-specialist applications, not just in language.

Going back to the meeting and the report, there is a large section on the need to foster these expert applications, for instance in health and biomedicine. Health is an enormous market and also a topic relevant for everyone in society. For instance, the report says that purpose-driven, domain-specialist systems "can outperform general-purpose models within their domains while requiring substantially less capital and compute," and calls this "a more achievable form of leadership than competing on raw scale" (§3.2).

The forum also identified a pressing need to build expertise in our institutions. AI literacy plays a large part in the delays that happen between bottom-up research and the corresponding responses by policy and institutions. We are starting a project with the German Federal Institute for Public Health that aims to describe the changing landscape of public health information and ultimately help the institutions to inform the population as best as possible in the post-GPT era.

Let's turn to SCIANCE. What is the Life Sciences working group, and what does it do?

If we see the Forum on Frontier AI as a top-down strategy of the Commission to drive their decisions from knowledge, the SCIANCE working groups are the corresponding bottom-up mechanism. Each of the ten working groups will work on strategic recommendations for AI-in-science questions in their topic for the next 2.5 years (informing the RAISE virtual institute). To achieve this, the organisers have established an asynchronous discourse forum and are planning workshops throughout the period for working group members to come together and discuss relevant topics. The first workshop I will participate in is on lab automation, co-located with the EGI conference in September.

Each working group will prioritise topics for informing European AI policy through the RAISE Secretariat. In the life sciences, we are expecting RAISE funding calls maybe in 2028; one thing I’d like to ensure is that those will be written with the best current knowledge about AI applications in the domain, and with the maximum benefit for the population in mind. 

If a colleague asked you about all this over lunch tomorrow, what key takeaways would you give them?

• The main failure mode is not the fast pace, nor is it powerlessness; it is paralysis.
• We would do well to focus on what we can do best, and invest most of our energy to make progress there (for instance, domain-specialist AI applications in biology and health).
• A common assumption is that most doors are locked; they are wide open, and most communities are very welcoming. If it matters enough to you, get involved.

Bottom line: fast and well-informed are not exclusive, nor a tradeoff. However, to push them both, we need to establish and hone trust between all members of any organisation.

How can Helmholtz AI further support scientists in the Helmholtz Association in alignment with that?

We should practice what we preach with respect to the request to the European Commission above: approach the problem top-down and bottom-up at the same time, find the right pace to make fast but well-informed decisions, and worry foremost about good communication in our organisation. The forum report emphasises the need for highest political priority and for leaders to stay “directly and continuously informed.” We are actively preparing instruments for keeping leadership informed, which should in time be scaled up to the “Paktforum” (AI working group of all German research institutes) and beyond.

This is why I believe we are on a good track. We formed an Agentic AI interest group more than a year ago already, which is transforming into a frontier AI task force as we speak. I am leveraging my experiences from a very similar task force I founded in ELIXIR (the European life-sciences data infrastructure). Now, with a new president, the Helmholtz Association is forming an “AI Action Plan,” which also includes such a task force. An initial outreach to colleagues across Helmholtz centers has just begun; our pre-existing task force should integrate into this effort. Helmholtz AI is the natural home for this task force and, to be pragmatic, should lead it actively (bottom-up) instead of observing and waiting for top-down to happen.

In my view, this task force ideally becomes a trusted platform for all Helmholtz members, from the scientists performing the groundwork to the highest tiers of management, who should enable and promote the scientists’ progress. This platform should become a communication channel where 1) experts come together regularly – say, monthly – to maintain a set of best practices for AI-in-science, and 2) governing bodies can get reliable, up-to-date information to base their decisions on in real time. If we manage to build that, with Helmholtz AI as the infrastructure provider, we will be able to make better decisions about AI in science, faster.