Europe added five ELLIS AI units. Make the network share more than a badge

THE SIGNAL IN ONE SENTENCE

A research network should be more than a handsome map covered in pins. On October 29, the European Laboratory for Learning and Intelligent Systems announced five new ELLIS Units in Côte d’Azur, Hamburg, Innsbruck, Ireland and Norway. The additions bring the network to 49 Units across 21 countries. Ireland and Norway are represented for the first time. That is a meaningful expansion. It is not yet a result. ELLIS describes its Units as local hubs that gather strong machine-learning researchers, connect institutions and develop long-term research agendas. The five newcomers cover a broad slice of the field. Côte d’Azur, based in Sophia Antipolis, is focused on foundations, health, digital infrastructure, creative industries and intelligent mobility. Its official profile also promises open access publications, open source repositories and open weights. Hamburg links five institutions around machine learning for natural sciences and engineering, including climate science, physics, biomedical science and cyber-physical systems. Innsbruck is interested in systems that can generalize beyond their training data and produce interpretable explanations. Ireland is organized as a national network that joins core machine learning with computer vision, language, information retrieval, health, the arts, humanities and social sciences. Norway connects seven participating institutions and puts particular emphasis on the Arctic and Nordic region, including climate, Earth observation, health, robotics and responsible AI. It is a very European collection of ambitions: climate models, medical systems, foundational work, public policy, the humanities, open research and several different definitions of trustworthy AI sitting at the same table. The table is the interesting part. ELLIS says every new Unit passed a formal selection process. Its current evaluation guidelines require an eligibility check, reviews from at least three international ELLIS Fellows, a selection-committee recommendation and a final board decision. The guidelines also ask whether a proposed Unit has enough researchers, sustainable funding, links to other ELLIS members and a credible contribution to the wider network. Those are serious entry conditions. They establish that the new groups are not random labs with a fresh sticker on the door. But selection evaluates a plan. A network earns its value after selection, when membership changes what researchers can actually do. That is why ELLIS should publish a shared-network receipt for every Unit. The receipt would not be a glossy annual report with four collaboration photos and a paragraph about excellence. It would be a structured, public record of the projects, exchanges, infrastructure, training and artifacts that moved through the network. Start with joint projects. For each project involving more than one Unit, list the participating teams, research question, start date, funding source, planned outputs and the specific contribution of the network. If the project would have happened anyway, say so. If ELLIS introduced the teams, unlocked a dataset or made a doctoral placement possible, record that too. The distinction matters because affiliation can be mistaken for causation. Two researchers may publish together while both happen to belong to ELLIS. That is useful science, but it is not automatically evidence that the network created the collaboration. Next, track people moving through the network. Count doctoral co-supervision, postdoctoral placements, visiting researchers, short exchanges, summer programs and staff shared across Units. Show the origin and destination, duration, discipline and whether travel or salary support was provided. Do not stop at a continental total. Break participation down by institution, career stage, gender where lawful and appropriate, and region. A network can look geographically broad while the same wealthy hubs receive most of the travel, compute and attention. Ireland and Norway joining for the first time makes this especially worth watching. Their researchers should gain access to the wider network, but the network should also absorb research questions shaped by Irish public institutions, Norwegian climate work and the realities of distributed research communities. Expansion is not just an invitation to join the existing conversation. It should change the conversation. Then show resource access. Compute is the obvious one. Record which Units can use shared clusters, accelerator allocations, cloud credits, secure environments and specialist hardware. Publish the application process, queue times, allocation size and completion rate. A researcher should be able to tell whether network membership provides practical access or merely another mailing list announcing that compute exists somewhere else. Data and facilities belong in the same ledger. Medical imaging collections, climate observations, robotics labs and public-sector datasets often have legal, ethical or technical restrictions. The receipt should show which resources can be discovered, which can be used across borders, what approvals are required and how long access takes. The system needs to distinguish availability from usability. A dataset listed in a portal is not shared if its license blocks the project. A cluster is not shared if only the host institution can run jobs. A visiting program is not accessible if researchers must cover housing in an expensive city without support. Open artifacts deserve their own section. ELLIS highlights open and reproducible research in the Côte d’Azur agenda. Turn that promise into fields: paper, code repository, model weights, dataset, evaluation harness, license, documentation, reproducibility instructions, compute requirements and maintenance status. If an artifact cannot be open, explain the boundary. Health data may need protection. Industry agreements may limit release. Security work may require staged access. Openness is not a magic word that makes every restriction irresponsible. It does, however, require specificity. The public should be able to see whether a project produced a usable benchmark or only a paper that says code will arrive later. Researchers should be able to find the model version used in an experiment and the environment needed to reproduce it. A repository that has not run since publication is an artifact, but it is not yet shared infrastructure. Training is another test of whether the network has substance. Each Unit says it wants to help educate the next generation. Publish the actual programs: courses, doctoral positions, co-supervision agreements, exchange seats, workshops, application dates, eligibility, language, funding and completion numbers. Make access visible to students outside the flagship institutions. The value of a European network is not that a student can admire excellent labs from another country. It is that the student can understand how to join a project, use a facility, find a supervisor or apply for a funded visit. Finally, record outcomes and the counterfactual. Outcomes can include publications, open tools, trained researchers, new companies, public-sector deployments, clinical evaluations, policy contributions and follow-on funding. Pair the numbers with short evidence notes. A citation count does not explain who benefited. A startup announcement does not show whether research moved into responsible use. The counterfactual is harder and more valuable: what happened because these groups were connected? Ask project leads to name the missing piece the network supplied. A partner with a complementary method. Access to Arctic observations. A secure medical environment. A second supervisor. A test site. A reviewer who caught a fragile assumption. A funded visit that turned a sketch into an experiment. This does not need to become a bureaucracy that steals time from research. The network already collects information for membership, funding, events and reporting. A common data model can let Units update one record and reuse it across the site, annual report and funder obligations. Keep the receipt small enough to maintain and detailed enough to audit. Every claim should link to evidence. Every total should expose its underlying projects. Every Unit should use the same definitions. Count an exchange when a person actually participates, not when a call opens. Count a shared artifact when another team can access it, not when a repository name is reserved. Count compute after allocation, not after a memorandum says collaboration is encouraged. The five new Units make a good pilot because they represent different shapes. Côte d’Azur and Hamburg are regional clusters inside countries that already have ELLIS Units. Innsbruck strengthens an existing Austrian ecosystem. Ireland is a national network crossing disciplines. Norway is geographically distributed across seven institutions. One template will not capture every local structure perfectly. That is fine. Use common core fields, then allow regional context around them. The network-wide view should answer four plain questions. Who can participate? What can they access? What did they make together? What would not have happened without the connection? ELLIS now has a wider map and more research communities inside it. The announcement establishes membership, ambitions and areas of expertise. It does not establish completed collaboration, equal access, open outputs or improved European competitiveness. Those are future claims. Give them receipts. If ELLIS can show people, resources and reusable work moving across its 49 Units, the network becomes infrastructure. If it can only show 49 names on a page, it remains a directory with excellent branding.