22 Academies, One Verdict: EASAC and FEAM Publish the Blueprint for Safe AI in European Healthcare

On 30 September 2026, at Arts 56 in Brussels and on livestream, two of Europe’s most consequential — and least flashy — scientific bodies publish their answer to one of the hardest questions in modern medicine: how do you bring artificial intelligence into healthcare without breaking it? The EASAC-FEAM joint report, formally titled AI in European Healthcare: Policy Recommendations for Optimising Added Value and Safe, Ethical, and Inclusive Adoption, is the product of a working group of 22 members drawn from academies across the continent, co-chaired by André Knottnerus for EASAC (the European Academies Science Advisory Council) and Luis Martí-Bonmatí for FEAM (the Federation of European Academies of Medicine). The collaboration has been running since 2024, and today its findings go public.

What the report actually covers

Strip away the launch-event pageantry — welcome remarks from EASAC Vice-President Birgitta Henriques-Normark and FEAM President Ferry Breedveld, key findings presented by FEAM Executive Director Louise Abboud and EASAC programme director Tom Cole-Hunter — and the report itself is a dense piece of policy engineering. It grapples with five interlocking problem domains:

Evidence generation. How do you prove an AI tool actually helps patients? The working group examines both randomised controlled trials — the gold standard that is often too slow and expensive for rapidly-iterating software — and pragmatic trial designs that trade some experimental purity for real-world clinical relevance. This is the central tension: a model that was validated on last year’s data may already be obsolete, yet a continuous-revalidation regime has no settled regulatory template anywhere in the world.

The regulatory thicket. A medical AI product in Europe today must navigate the Medical Device Regulation (MDR), the AI Act, the European Health Data Space (EHDS), and the Product Liability Directive — four legislative frameworks that were written largely independently of one another. The report maps where they align and, more importantly, where they conflict or leave gaps.

Data interoperability. Europe’s health systems are famously fragmented: 27 member states, dozens of electronic health record standards, and uneven digital maturity. An AI model trained on data from one system may quietly fail on another. The report treats interoperability not as an IT afterthought but as a precondition for safe deployment.

Ethics, privacy, and explainability. Transparency, model explainability, equitable access, and privacy get their own treatment. The explainability question is particularly sharp in medicine: a clinician who cannot understand why a model recommends against a treatment cannot meaningfully consent to using it, yet the most performant models are often the least interpretable.

Workforce and education. Perhaps the least glamorous and most overlooked section. AI tools fail in hospitals not because the algorithms are wrong but because the radiologists, nurses, and administrators around them were never trained to use, question, or override them. The working group argues that education reform is a first-order deployment requirement, not a nice-to-have.

Who is in the room

The launch programme is itself a snapshot of who owns this problem in Europe. Panel I, on “Regulation, validation and accountability,” features Saila Rinne, head of the AI in Health and Life Sciences unit at the EU AI Office (DG CNECT), and Andreia M. Oliveira, policy officer for AI in healthcare at DG SANTE — in other words, the two European Commission directorates that will have to act on whatever the report recommends. They sit alongside working group members Miikka Korja (a neurosurgeon) and Eva Fialová, moderated by Martí-Bonmatí.

Panel II turns to “Implementation, trust and equity” — what it takes to deploy AI safely inside a hospital and what it takes for patients to trust it once it is there. Speakers include Sascha Marschang of the European Hospital and Healthcare Federation (HOPE), Bianca Pop of the European Patients’ Forum (EPF), and working group members Christian Lovis and David Ríos Insua, moderated by Knottnerus. The presence of both the hospital federation and the patients’ forum signals the report’s core stance: regulation written without implementers and patients at the table produces rules that nobody follows.

Why this matters beyond Brussels

It is tempting to file this under “yet another EU policy paper.” Three things make it more than that.

First, the timing. The AI Act’s obligations for high-risk systems — and the Act explicitly classifies most healthcare AI as high-risk — are phasing in right now, while the EHDS regulation, the EU’s first common data space, is moving from adopted text to operational reality. The MDR conformity-assessment backlog is already a scandal in medtech. Recommendations landing in this window can actually shape implementing guidance rather than gather dust.

Second, the provenance. EASAC and FEAM are academy networks — the national science academies and national medical academies of essentially all of Europe, operating within the SAPEA ecosystem that formally advises the European Commission’s Scientific Advice Mechanism. When these bodies converge on a position, it carries the weight of the scientific establishment rather than of any single lobby. Their 22-member working group took two years to get here.

Third, the substance. The hardest question the report confronts is accountability: when an AI-supported clinical decision harms a patient, who answers? The updated Product Liability Directive now covers software, but the chain from model developer to hospital deployer to treating clinician is long and each link points at the others. The report’s answer — that validation, monitoring, and clear accountability allocation must be designed in before deployment, not litigated after — is the difference between trust and backlash.

The road from here

A report launch is a beginning, not an end. The realistic best case is that the report’s recommendations feed into the Commission’s implementation guidance for the AI Act’s high-risk provisions in health, into EHDS secondary-use rules for model training and validation, and into national procurement standards that hospitals can cite when vendors come knocking. The pessimistic case is that the pace of commercial AI deployment outruns all of it, and the report becomes a post-mortem document.

What the working group has produced, at minimum, is the most complete map to date of the terrain: where the evidence standards should sit, which laws collide, whose hands must be on the wheel, and what patients are owed. In a domain where the default trajectory is hype-then-scandal, a sober two-year academy effort that lands before the mass deployment wave is exactly the kind of boring, vital work that rarely makes headlines — and exactly the kind that determines whether AI in European medicine ends up trusted or tolerated.