The EU AI Act needs
Foundation Model Regulation
To the members of the German Federal Government
We are a group of German and international experts in the field of AI and leaders in business, civil society, and academia. Our expertise is diverse and significant: among us are the two most cited AI researchers in the world and recipients of the Turing Award (“Nobel Prize of Computing”), the author of the world’s most widely used textbook on AI, experts who have advised the German Federal Government on AI and adjacent topics, and founders of successful AI companies.
We understand the European Union’s AI Act is nearing finalization and is presently being discussed in trilogue. In our shared view, the AI Act could become a landmark work of legislation, shaping the future of artificial intelligence not only in the EU, but the entire world. We commend the EU and its member states for their serious and appropriate treatment of this important issue, and their global leadership on it.
In our view, however, the Act’s potential is now at risk: a central element of the AI Act, namely binding rules for foundation models, faces resistance from some member states. We understand that the German Federal Government is part of this opposition.
We believe regulating foundation models in the AI Act is vital for a flourishing and safe AI ecosystem. We strongly advise against dealing with foundation models merely through a system of self-regulation.
Many of the world’s most esteemed AI experts have been cautioning against the manifold risks from advanced AI, including risks to public safety such as AI-driven disinformation and manipulation, AI-enhanced cyber attacks, or AI-generated pathogens. Such risks arise predominantly from the most powerful foundation models, as reflected in the White House’s recent Executive Order on AI and this month’s historic Bletchley Declaration, signed by 28 countries and the EU. These risks to public safety are inherent to foundation models. Therefore, they should be addressed at the foundation model level. The provisions for foundation models envisioned by the EU Parliament and the Spanish Council presidency – relating to e.g. safety and cybersecurity, risk assessment and mitigation systems, pre-deployment red-teaming, and post-deployment auditing – are therefore essential for a flourishing and safe AI ecosystem in the EU because they address such inherent risks.
Such binding rules are important for both economic and safety reasons. Economically, ensuring the safety of foundation models is a necessity for thousands of SMEs and other downstream deployers who want to use these foundation models for their innovative products. They cannot afford liability risks and excessive compliance costs stemming from a potentially unsafe foundation model they use for their product. Exempting foundation models from the AI Act would therefore severely stifle innovation.
From a safety perspective, too, it is vital that risks are addressed at the foundation model level. Only the providers of foundation models are in a position to comprehensively address their inherent risks. They exclusively have access to and knowledge of the models’ training data, guardrail design, likely vulnerabilities, and other core properties. If severe risks from foundation models aren’t mitigated at the foundation model level, they won’t be mitigated at all, potentially threatening the safety of millions of people.
We understand some voices support addressing risks from foundation models through a system of self-regulation. We strongly advise against this. Self-regulation is likely to dramatically fall short of the standards required for foundation model safety. Since even a single unsafe model could cause risks to public safety, a vulnerable consensus on self-regulation does not ensure EU citizens’ safety. The safety of foundation models must be ensured by law.
The AI Act, if it includes foundation models, would be the world’s first comprehensive regulation of AI, viewed as a historic example of European leadership. If coverage of foundation models is dropped, a weakened or failed AI Act would be regarded as a historic failure.
We therefore strongly encourage the German Federal Government to leverage its leadership in the European Union to ensure the inclusion of comprehensive foundation model regulation in the EU AI Act.
Signatories
Prof. em. Geoffrey Hinton, University of Toronto, Chief Scientific Adviser at the Vector Institute, 2018 Turing Award Winner
Prof. Yoshua Bengio, Université de Montréal, Founder and Scientific Director of Mila – Quebec AI Institute, 2018 Turing Award Winner
Prof. em. Gary Marcus, NYU, Founder and CEO, Geometric Intelligence (acquired by Uber)
Prof. Stuart Russell, UC Berkeley, Director of the Center for Human-Compatible Artificial Intelligence, co-author of the standard textbook “Artificial Intelligence: a Modern Approach"
Marietje Schaake, International Policy Fellow, Stanford Institute for Human-Centered Artificial Intelligence
Andreas Loy, Founder & CEO, KONUX
Prof. Holger Hoos, RWTH Aachen University & University of Leiden
Prof. em. Raja Chatila, Sorbonne University
Prof. Dr. jur. Silja Vöneky, Universität Freiburg
Prof. Karl Hans Bläsius, Hochschule Trier
Prof. Wolfgang Schröder, Julius-Maximilians-Universität Würzburg
Prof. Christoph Benzmüller, Chair for AI Systems Development, Otto-Friedrich-Universität Bamberg
Prof. Gerhard Lakemeyer, Chair, Department of Computer Science, RWTH Aachen
Prof. Otthein Herzog, Universität Bremen
Prof. Mathias Risse, Harvard University, Director of the Carr Center for Human Rights Policy
Prof. Marius Lindauer, Leibniz Universität Hannover
Prof. Katharina Morik, TU Dortmund, AI Chair (emerita)
Prof. Wil van der Aalst, RWTH Aachen
Kaltrina Shala LL.M., LL.M., Weizenbaum-Institut e.V.
Prof. Peter Struss, TU Munich
Prof. Kai-Uwe Kühnberger, University Professor for Artificial Intelligence, Osnabrück University
Prof. Dr. Claus Rollinger, Universität Osnabrück
Prof. Amparo Lasen, Universidad Complutense de Madrid
Prof. Henny van der Windt, Associate Professor Science and Technology Studies and Environment, University of Groningen
Prof. Elisabeth Wesseling, Maastricht University
Prof. Karsten Weber, Professor for Technology Assessment and AI-based Mobility, OTH Regensburg
Prof. Alessandro Caliandro, Università degli Studi di Pavia
Prof. Alex Gekker, Assistant Professor in Digital Research Methods, Universitiy of Amsterdam
Prof. Dino Pedreschi, Member of the Scientific Board of the EU program “FAIR - Future AI Research”, Italian delegate in the Global Partnership on AI, University of Pisa
Prof. Mykola Pechenizkiy, TU Eindhoven
Prof. Tim Kietzmann, Professor for Machine Learning, Universität Osnabrück
Prof. Dr. Abdur Razzaque Khan, University of Dhaka
Prof. Estrid Sørensen, Ruhr-University Bochum
Marc Rotenberg, Executive Director, Center for AI and Digital Policy
Prof. Cordula Kropp, Universität Stuttgart
Univ.-Prof. Hannes Werthner, TU Wien
Prof. Tapabrata Rohan Chakraborty, Honorary Associate Professor in Transparent AI, University College London, Senior Research Associate at the Alan Turing Institute, invited expert in Responsible AI with the Global Partnership on AI
Prof. Martin Butz, University of Tübingen
Prof. Jefrey Lijffijt, Professor of Data Science, Knowledge Discovery, and Visual Analytics, Ghent University
Prof. Sven Koenig, University of Southern California
Prof. Andreas Weber, University of Twente
Gilles Moyse, CEO, reciTAL
Alistair Knott, Co-lead, Responsible AI for Social Media Governance, Global Partnership on AI
Sharon Polsky, President of the Privacy & Access Council of Canada
Prof. Gilles Escarguel, Associate Professor in Macroecology, Université Lyon 1
Prof. Peter Thompson, Victoria University of Wellington
Prof. Tobias Matzner, Professor for Digital Humanities, Paderborn University
Prof. Volker Brühl, Professor of Banking and Finance, Goethe University
Annika Brack, CEO, International Center for Future Generations
Alessandra Sala, President, Women in AI
Prof. Cinzia Padovani, Southern Illinois University
Prof. Peter König, University Osnabrück
Prof. Luciano Floridi, Director, Yale Digital Ethics Center
Prof. Massimiliano Simons, Assistant Professor in Philosophy of Technology, Maastricht University
Prof. Bruno Caldas Pires, Universitat Politècnica de Catalunya
Maciej Chojnowski, Co-Founder & Program Director, Center for Ethics of Technology at the Humanities Institute
Esther Hammelburg, Senior Lecturer, Amsterdam University of Applied Sciences
Prof. Ronald Leenes, Professor of Regulation by Technology & Former Director, Tilburg Institute for Law, Technology, and Society
Rufo Guerreschi, President, Trustless Computing Association
Prof. Lina Eklund, Uppsala University
Dr. Eleanor O'Leary, Lecturer in Media and Communications, South East Technological University
Prof. Thomas Potthast, Professor and Director, Centre for Ethics in the Sciences and Humanities, University of Tübingen
Prof. Marc Pananceau, Paris-Saclay University
Dr. Susan Leavy, University College Dublin, Irish Delegate at the Global Partnership on AI
Prof. Wouter Boon, Utrecht University
Domenico Fiormonte, Lecturer in the Sociology of Communication and Culture, University of Roma Tre
Prof. Olya Kudina, Assistant Professor AI Ethics, TU Delft
Prof. James Steinhoff, University College Dublin
Prof. Maciej Piasecki, Wroclaw University of Science and Technology
Prof. Dr. Ingrid Schneider, Universität Hamburg
Prof. Felix Wichmann, Universität Tübingen
Ermelinda Kanushi, Governance Lecturer, University College Freiburg
Przemyslaw Barchan, Director, Institute of LegalTech of the National Bar Council (Poland)
Prof. Lisa McLaughlin, Associate Professor, Miami University
Prof. Federico Faroldi, Professor of Ethics, Law and AI, University of Pavia; Director, Normative Risk Lab
Theresa Züger, Head of AI and Society Lab, Humboldt Institut für Internet und Gesellschaft
Prof. Christian Herzog, Professor of the Ethical, Legal and Social Aspects of AI, University of Lübeck