From July 13 to 19, Hi! PARIS researchers will take part in the International Conference on Machine Learning (ICML 2025), one of the world’s leading conferences in artificial intelligence, held this year in Vancouver, Canada. A total of 27 research papers from Hi! PARIS affiliated teams have been accepted following ICML’s highly selective peer-review process, recognizing the excellence and innovation of our researchers across partner institutions. This strong presence reflects Hi! PARIS’s scientific leadership and deep engagement in frontier AI research. The accepted papers span a wide array of domains, including machine learning theory, optimization, generative models, privacy, and multi-agent systems, and demonstrate the center’s interdisciplinary strength. By contributing to ICML 2025’s core sessions and workshops, Hi! PARIS reaffirms its commitment to advancing AI research for science, business, and society at the highest international level. Congratulations to our researchers! Here is a list of papers accepted at ICML 2025 that include at least one author affiliated with Hi! PARIS: Title Hi! PARIS Authors All Authors Branches: Efficiently Seeking Optimal Sparse Decision Trees via AO* Albert Bifet, Jesse Read Ayman Chaouki, Jesse Read, Albert Bifet To Each Metric Its Decoding Thomas Bonald, Matthieu Labeau Roman Plaud, Alexandre Perez-Lebel, Matthieu Labeau, Antoine Saillenfest, Thomas Bonald Scaling Laws for Forgetting Marco Cuturi Louis Béthune, David Grangier, Dan Busbridge, Eleonora Gualdoni, Marco Cuturi, Pierre Ablin Shielded Diffusion Marco Cuturi Michael Kirchhof, James Thornton, Louis Béthune, Pierre Ablin, Eugene Ndiaye, Marco Cuturi Misspecification in Simulation-based Inference Marco Cuturi Antoine Wehenkel, Juan L. Gamella, Ozan Sener, Jens Behrmann, Guillermo Sapiro, Jörn Jacobsen, Marco Cuturi TabICL Gael Varoquaux Jingang QU, David Holzmüller, Gael Varoquaux, Marine Le Morvan Byzantine Robust Gossip Aymeric Dieuleveut Renaud Gaucher, Aymeric Dieuleveut, Hadrien Hendrikx Scaffold with Stochastic Gradients Aymeric Dieuleveut, Alain Oliviero Durmus, Eric Moulines Paul Mangold, Alain Oliviero Durmus, Aymeric Dieuleveut, Eric Moulines Compressed and Distributed Least-Squares Aymeric Dieuleveut Constantin Philippenko, Aymeric Dieuleveut Discrete Markov Probabilistic Models Alain Oliviero Durmus Le Tuyet Nhi PHAM, Dario Shariatian, Antonio Ocello, Giovanni Conforti, Alain Oliviero Durmus Prediction-Aware Learning Alain Oliviero Durmus, Eric Moulines Aymeric Capitaine, Etienne Boursier, Eric Moulines, Michael Jordan, Alain Oliviero Durmus Mixture-Based Framework for Diffusion Models Alain Oliviero Durmus, Eric Moulines Yazid Janati el idrissi, Badr MOUFAD, Mehdi Qassime, Alain Oliviero Durmus, Eric Moulines, Jimmy Olsson Differential Privacy for MCMC Alain Oliviero Durmus Andrea Bertazzi, Tim Johnston, Gareth Roberts, Alain Oliviero Durmus Asymmetric Actor-Critic Algorithms Damien Ernst Gaspard Lambrechts, Damien Ernst, Aditya Mahajan Score-Based Generative Models in W2 Marta Gentiloni Silveri Marta Gentiloni Silveri, Antonio Ocello Sliced-Wasserstein Distance Analysis Anna Korba Christophe Vauthier, Anna Korba, Quentin Mérigot Density Ratio Estimation Anna Korba Hanlin Yu, Arto Klami, Aapo Hyvarinen, Anna Korba, Lemir Omar Chehab Wasserstein Gradient Flows Anna Korba Clément Bonet, Christophe Vauthier, Anna Korba Learning of Continuous Markov Semigroups Karim Lounici Vladimir Kostic, Karim Lounici, Hélène Halconruy, Timothée Devergne, Pietro Novelli, Massimiliano Pontil GNN with GMM Augmentation Johannes Lutzeyer, Michalis Vazirgiannis Yassine Abbahaddou, Fragkiskos Malliaros, Johannes Lutzeyer, Amine Aboussalah, Michalis Vazirgiannis TRPO in Mean Field Games Eric Moulines Antonio Ocello, Daniil Tiapkin, Lorenzo Mancini, Mathieu Lauriere, Eric Moulines Conditional Coverage with Conformity Scores Eric Moulines Vincent Plassier, Alexander Fishkov, Victor Dheur, Mohsen Guizani, Souhaib Ben Taieb, Maxim Panov, Eric Moulines Efficient On-Device Learning Van-tam Nguyen, Enzo Tartaglione Le-Trung Nguyen, Aël Quélennec, Van-Tam Nguyen, Enzo Tartaglione Pareto-Optimality in One-Max-Search Vianney Perchet Ziyad Benomar, Lorenzo Croissant, Vianney Perchet, Spyros Angelopoulos Last Iterate Convergence for Uncoupled Learning Vianney Perchet Côme Fiegel, Pierre Menard, Tadashi Kozuno, Michal Valko, Vianney Perchet Quantifying Treatment Effects Erwan Scornet Ahmed Boughdiri, Julie Josse, Erwan Scornet Prediction via Shapley Value Regression Michalis Vazirgiannis Amr Alkhatib, Roman Bresson, Henrik Boström, Michalis Vazirgiannis
We are proud to introduce the newest cohort of Hi! PARIS Chairs, part of the Hi! PARIS Cluster 2030 initiative. Designed to foster long-term excellence in AI and data analytics, this program supports innovative research across science, business, and society. Hosted within leading institutions, Institut Polytechnique de Paris, HEC Paris, Inria, and CNRS, the 2025 chairs were selected through a highly competitive open call. Our new cohort includes 19 exceptional researchers, spanning Standard (Starting & Advanced) and Synergy Fellowships. They are hosted across Hi! PARIS partner instituions and bring diverse, forward-thinking AI applications: Starting Fellowships Gül Varol (ENPC) – Training Machines to Train Humans in Motion Enzo Tartaglione (Télécom Paris) – Frugal AI: Grounding Information Flow Vicky Kalogeiton (Ecole polytechnique) – Scalable Model-Data Co-evolution Stephan Alaniz (Télécom Paris) – Explaining & Removing Social Biases in Text-to-Image AI Amaury Hayat (ENPC) – Efficient Symbolic Mathematics Systems Advanced Fellowships Antonin Bergeaud (HEC Paris) – AI Adoption: Economic & Social Implications Loïc Landrieu (ENPC) – Universal Geospatial Representations Ioana Manolescu (Inria) – AI Data Dialogs for the Press Alain Oliviero-Durmus (Ecole polytechnique) – Enhanced Generative Models Etienne Ollion (CNRS) – Textual Politics in Traditional Media Fabian Suchanek (Télécom Paris) – Lossy Language Models Synergy Fellowships Florence D’Alché-Buc (Télécom Paris), Rémi Flamary (Ecole polytechnique), Charlotte Laclau (Télécom Paris) & Karim Lounici (Ecole polytechnique) – Advancing efficient, reliable and science-informed Learning for non-euclidean data with Application to molecule and biological network Structures Patrick Loiseau (Inria), Vianney Perchet (ENSAE Paris), Yuki Tamura (Ecole polytechnique) & Pablo Winant (Ecole polytechnique) – Design, Incentivization, Optimization and Reinforcement Learning of Multi-Layered Market. This fellowship cohort reflects the depth and diversity of the Hi! PARIS research ecosystem. From geospatial modeling and symbolic reasoning to AI ethics, media analysis, and market dynamics, the selected projects embody the center’s commitment to advancing interdisciplinary and socially impactful research. Through its Starting, Advanced, and Synergy Fellowship tracks, Hi! PARIS is proud to support both emerging talent and established leaders, fostering collaboration across institutions and disciplines to drive the future of AI for science, business, and society.
Exceptional Seminar with Prof. Eric Xing at École polytechnique Hi! PARIS is honored to welcome Professor Eric Xing, President of the Mohamed bin Zayed University of Artificial Intelligence and Professor at Carnegie Mellon University, for an exceptional scientific seminar on July 8, 2025, at École Polytechnique. As one of the leading figures in AI and machine learning, Prof. Xing will present his latest work on general and purposeful reasoning systems beyond traditional language models. This event is part of our commitment to fostering high-level scientific exchange and advancing cutting-edge AI research. Abstract The success of large language models (LLMs) like OpenAI’s GPTs has amazed the world with their extraordinary performance in tasks such as standardized tests, advanced mathematical reasoning, encyclopedic knowledge retrieval, and human-like language generation. However, these models fall short when it comes to embodied reasoning, physical and social understanding, and real-world strategic planning. In this seminar, Prof. Xing will introduce “PAN”, a novel architecture designed for general and purposeful reasoning in complex, real-world environments. PAN combines a new World Model for steerable simulations of potential outcomes and a dynamic Agent Model capable of planning, learning, and adapting to achieve goals across diverse contexts. He will contrast this approach with current architectures, addressing common misconceptions around agent and world models, and discuss choices of data, representations, and training strategies. The talk will also feature theoretical insights and a preview of an upcoming model release. Prof. Xing will conclude with a broader reflection on concepts such as agency and artificial general intelligence (AGI), informed by a renewed philosophical lens. About the Speaker Professor Eric Xing is the President of the Mohamed bin Zayed University of Artificial Intelligence and a Professor of Computer Science at Carnegie Mellon University. His research spans machine learning, large-scale distributed computing, and statistical modeling, with recent focus on foundational models in biology and AI. He has received multiple prestigious awards, including the NSF CAREER Award, the Sloan Fellowship, and best paper awards at leading conferences like ACL, ISMB, NeurIPS, and OSDI. He is a Fellow of the AAAI, ACM, ASA, IEEE, and IMS, and has served on editorial boards for top journals in the field.
We are proud to announce that Arnak Dalalyan, director of CREST, Hi! PARIS Fellow and Professor of statistics at ENSAE Paris, has been awarded a European Research Council (ERC) Advanced Grant for his project SAGMOS.
Télécom Paris PhD researcher Ali Keshavarzi reveals how AI can model the lungs more accurately, using fewer annotated scans, by combining deep learning with frugal, topology-aware techniques.
Hi! PARIS launched the International Visiting Chairs program in December 2024 to support international scientific collaboration. The program invites researchers from institutions abroad to spend time working with research teams affiliated with Hi! PARIS member schools in France. For 2025, we are welcoming seven international visiting professors. Each will collaborate on a specific project in artificial intelligence or data science, hosted by a Hi! PARIS researcher. Meet the 2025 Visiting Chairs: Alessandro VinciarelliProject: Socially Intelligent Multimodal Conversation AnalysisHost: Mounim El Yacoubi, Télécom SudParis Alexei EfrosProject: Discovering Typical Visual Structures in Large Image Collections with Diffusion ModelsHost: Mathieu Aubry, ENPC Andréas NüchterProject: Universal Semantic Mapping for Outdoor and Underwater EnvironmentsHost: François Goulette, ENSTA Dimitris SamarasProject: Human Attention-Guided Video RepresentationsHost: Vicky Kalogeiton, École polytechnique Eric XingProject: Multiscale Foundation Models for Predicting, Simulating, and Programming Biology at All LevelsHosts: Eric Moulines & Alain Durmus, École polytechnique Milan MiricProject: Identifying US, Chinese, and European AI Technologies: Extending Patent Databases Using Novel Data and ApproachesHost: Mickael Impink, HEC Paris Thibaut VidalProject: Trustworthy and Decision-Focused LearningHost: Axel Parmentier, ENPC This initiative reflects our core values of openness, excellence, and cross-border collaboration. By welcoming top researchers into our ecosystem, we aim to accelerate innovation and contribute to advance the global conversation around responsible and impactful AI.
In April 2025, Hi! PARIS researchers were recognized at two major international conferences in Artificial Intelligence. From April 6 to 11, the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) highlighted breakthroughs in signal processing, audio, and speech technologies. Later, from April 24 to 28, the International Conference on Learning Representations (ICLR), the premier gathering for experts advancing the field of representation learning brought together leading minds in machine learning and deep learning. Across both events, 41 research papers by Hi! PARIS-affiliated teams were accepted, underscoring the center’s commitment to cutting-edge, interdisciplinary research in AI and data science. Congratulations to our researchers! List of papers accepted at ICASSP and ICLR 2025 Conference Title Hi! PARIS Authors All Authors ICLR 2025 Probabilistic Conformal Prediction with Approximate Conditional Validity Eric Moulines Vincent Plassier, Alexander Fishkov, Mohsen Guizani, Maxim Panov, Eric Moulines ICLR 2025 From Risk to Uncertainty: Generating Predictive Uncertainty Measures via Bayesian Estimation Eric Moulines Nikita Kotelevskii, Vladimir Kondratyev, Martin Takáč, Eric Moulines, Maxim Panov ICLR 2025 Variational Diffusion Posterior Sampling with Midpoint Guidance Yazid Janati el idrissi, Lisa Bedin, Alain Oliviero Durmus, randal douc, Eric Moulines Badr MOUFAD, Yazid Janati el idrissi, Lisa Bedin, Alain Oliviero Durmus, randal douc, Eric Moulines, Jimmy Olsson ICLR 2025 Nonasymptotic Analysis of Stochastic Gradient Descent with the Richardson–Romberg Extrapolation Alain Oliviero Durmus, Eric Moulines Marina Sheshukova, Denis Belomestny, Alain Oliviero Durmus, Eric Moulines, Aleksei Naumov, Sergey Samsonov ICLR 2025 Imputation for prediction: beware of diminishing returns Gael Varoquaux Marine Le Morvan, Gael Varoquaux ICLR 2025 Learned Reference-based Diffusion Sampler for Multi-Modal Distributions Alain Oliviero Durmus Maxence Noble, Louis Grenioux, Marylou Gabrié, Alain Oliviero Durmus ICLR 2025 Denoising Levy Probabilistic Models Alain Oliviero Durmus Dario Shariatian, Umut Simsekli, Alain Oliviero Durmus ICLR 2025 Watermark Anything With Localized Messages Alain Oliviero Durmus Tom Sander, Pierre Fernandez, Alain Oliviero Durmus, Teddy Furon, Matthijs Douze ICLR 2025 Building Blocks of Differentially Private Training Aymeric Dieuleveut Mahmoud Hegazy, Aymeric Dieuleveut ICLR 2025 Tailoring Mixup to Data for Calibration Florence d’Alché-Buc Quentin Bouniot, Pavlo Mozharovskyi, Florence d’Alché-Buc ICLR 2025 Restyling Unsupervised Concept Based Interpretable Networks with Generative Models Florence d’Alché-Buc Jayneel Parekh, Quentin Bouniot, Pavlo Mozharovskyi, Alasdair Newson, Florence d’Alché-Buc ICLR 2025 Long-time Asymptotics of Noisy SVGD Outside the Population Limit Pascal Bianchi Victor Priser, Pascal Bianchi, Adil Salim ICLR 2025 Solving Differential Equations with Constrained Learning Luiz Chamon Viggo Moro, Luiz Chamon ICLR 2025 Simple ReFlow: Improved Techniques for Fast Flow Models Marco Cuturi Beomsu Kim, Yu-Guan Hsieh, Michal Klein, Marco Cuturi, Jong Chul YE, Bahjat Kawar, James Thornton ICLR 2025 Controlling Language and Diffusion Models by Transporting Activations Marco Cuturi Pau Rodriguez, Arno Blaas, Michal Klein, Luca Zappella, Nicholas Apostoloff, Marco Cuturi, Xavier Suau ICLR 2025 Disentangled Representation Learning with the Gromov-Monge Gap Marco Cuturi Théo Uscidda, Luca Eyring, Karsten Roth, Fabian Theis, Zeynep Akata, Marco Cuturi ICLR 2025 An Illustrated Guide to Automatic Sparse Differentiation Guillaume Dalle Adrian Hill, Guillaume Dalle, Alexis Montoison ICLR 2025 Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It Gianni Franchi Guoxuan Xia, Olivier Laurent, Gianni Franchi, Christos-Savvas Bouganis ICLR 2025 Last-Iterate Convergence Properties of Regret-Matching Algorithms in Games Julien Grand-Clément Yang Cai, Gabriele Farina, Julien Grand-Clément, Christian Kroer, Chung-Wei Lee, Haipeng Luo, Weiqiang Zheng ICLR 2025 Provable Convergence and Limitations of Geometric Tempering for Langevin Dynamics Anna Korba, Austin Stromme Omar Chehab, Anna Korba, Austin Stromme, Adrien Vacher ICLR 2025 NextBestPath: Efficient 3D Mapping of Unseen Environments Vincent Lepetit Shiyao Li, Antoine Guedon, Clémentin Boittiaux, Shizhe Chen, Vincent Lepetit ICLR 2025 Understanding Virtual Nodes: Oversquashing and Node Heterogeneity Johannes Lutzeyer Joshua Southern, Francesco Di Giovanni, Michael Bronstein, Johannes Lutzeyer ICLR 2025 Restyling Unsupervised Concept Based Interpretable Networks with Generative Models Pavlo Mozharovskyi, Florence d’Alché-Buc Jayneel Parekh, Quentin Bouniot, Pavlo Mozharovskyi, Alasdair Newson, Florence d’Alché-Buc ICLR 2025 Tailoring Mixup to Data for Calibration Pavlo Mozharovskyi, Florence d’Alché-Buc Quentin Bouniot, Pavlo Mozharovskyi, Florence d’Alché-Buc ICLR 2025 AtomSurf: Surface Representation for Learning on Protein Structures Maks Ovsjanikov Vincent Mallet, Yangyang Miao, Souhaib Attaiki, Bruno Correia, Maks Ovsjanikov ICLR 2025 Feature-Based Online Bilateral Trade Vianney Perchet Solenne Gaucher, Martino Bernasconi, Matteo Castiglioni, Andrea Celli, Vianney Perchet Conference Title Hi! PARIS Authors All Authors ICASSP 2025 A Hybrid Model for Weakly-Supervised Speech Dereverberation Mathieu Fontaine, Gaël Richard Louis Bahrman, Mathieu Fontaine, Gaël Richard ICASSP 2025 AnCoGen: Analysis, Control and Generation of Speech with a Masked Autoencoder Gaël Richard Samir Sadok, Simon Leglaive, Laurent Girin, Gaël Richard, Xavier Alameda-Pineda ICASSP 2025 Contrastive Knowledge Distillation for Embedding Refinement in Personalized Speech Enhancement Mathieu Fontaine, Slim Essid Thomas Serre, Mathieu Fontaine, Éric Benhaim, Slim Essid ICASSP 2025 F-STRIPE: Fast Structure-Informed Positional Encoding for Symbolic Music Generation Gaël Richard Manvi Agarwal, Changhong Wang, Gaël Richard ICASSP 2025 Investigating the Sensitivity of Pre-trained Audio Embeddings to Common Effects Gaël Richard Victor Deng, Changhong Wang, Gaël Richard, Brian McFee ICASSP 2025 Learning Source Disentanglement in Neural Audio Codec Gaël Richard Xiaoyu Bie, Xubo Liu, Gaël Richard ICASSP 2025 Masked Latent Prediction and Classification for Self-Supervised Audio Representation Learning Geoffroy Peeters, Slim Essid Aurian Quelennec, Pierre Chouteau, Geoffroy Peeters, Slim Essid ICASSP 2025 Multiple Choice Learning for Efficient Speech Separation with Many Speakers David Perera, Gaël Richard, Slim Essid David Perera, François Derrida, Théo Mariotte, Gaël Richard, Slim Essid ICASSP 2025 O-EENC-SD: Efficient Online End-to-End Neural Clustering for Speaker Diarization Mathieu Fontaine, Slim Essid Elio Gruttadauria, Mathieu Fontaine, Jonathan Le Roux, Slim Essid ICASSP 2025 Perceptual Noise-Masking with Music through Deep Spectral Envelope Shaping Roland Badeau, Slim Essid Clémentine Berger, Roland Badeau, Slim Essid ICASSP 2025 Re-Evaluating Privacy in Centralized and Decentralized Learning: An Information-Theoretical and Empirical Study Changlong Ji, Stephane Maag Changlong Ji, Richard Heusdens, Stephane Maag, Qiongxiu Li ICASSP 2025 Standardization Status of MPEG Video-based Dynamic Mesh Coding (V-DMC) Marius Preda Wenjie Zou, Shizhuo Zhang, Fuzheng Yang, Marius Preda ICASSP 2025 Targeted Data Poisoning for Black-Box Audio Datasets Ownership Verification El-Mahdi El-Mhamdi Wassim Wes Bouaziz, El-Mahdi El-Mhamdi, Nicolas Usunier ICASSP 2025 Twenty-Five Years of MIR Research: Achievements, Practices, Evaluations, and Future Challenges Geoffroy Peeters
In February 2025, nearly five years after the creation of the Center, we hosted a dinner to thank our corporate donors for their continued support and commitment. It was an opportunity to reflect on the progress we’ve made together and to reaffirm the importance of the partnerships that have helped shape Hi! PARIS into what it is today. When Hi! PARIS was launched in 2020 by Institut Polytechnique de Paris and HEC Paris, and later joined by Inria, CNRS, and UTT, the goal was clear: to build an ambitious and internationally visible center for research, education, and innovation in artificial intelligence and data science. From the beginning, our corporate donors have played a central role in supporting this vision, not only financially, but also intellectually, by sharing their priorities and contributing to the development of our initiatives, research agenda, and educational programs. Pictured from left to right: Éloïc Peyrache, Dean of HEC Paris; Thierry Coulhon, President of Institut Polytechnique de Paris; Philippe Baptiste, French Minister for Research and Higher Education and Michael I. Jordan, Professor at UC Berkeley. Over the past five years, these partnerships have enabled us to launch and fund research projects at the crossroads of science, business, and society; support chair holders, fellows, postdoctoral researchers, and PhD students working on key AI challenges; host events that connect academic and industrial communities; and develop interdisciplinary teaching programs to train the next generation of AI and data talent. These achievements would not have been possible without the trust and involvement of our donors. Together, we’ve helped build a center that stands for scientific excellence, responsible innovation, and collaboration across sectors. https://www.youtube.com/watch?v=CrEHR_vHhcg A committed community of donors The strength of Hi! PARIS lies not only in its academic and scientific foundations, but also in the commitment of its corporate donors. Partners such as L’Oréal, Capgemini, TotalEnergies, VINCI, and Schneider Electric bring a diverse range of perspectives from industry, technology, and energy to infrastructure and innovation united by a shared belief: that artificial intelligence and data science must be developed in a way that serves society. Each of them brings a unique perspective, and their support goes far beyond funding. They contribute ideas, challenge us to stay relevant, and help us anchor our work in the real world. This richness of backgrounds and their shared values of excellence, responsibility, and openness reinforce Hi! PARIS’s mission to bridge science, business, and society. One of the most concrete illustrations of this collaboration is our white book Visions of Business: Driving Business Innovation with Data & AI. This 48-page publication brings together real-world use cases from our corporate donors and highlights how they are leveraging AI to address key challenges in beauty, energy, retail, infrastructure, and beyond. Pictured from left to right: Jean-Paul Agon, Chairman of the Board at L’Oréal; Laurent Bataille, President of Schneider Electric France; Philippe Rambach, Chief AI Officer at Schneider Electric; Delphine Colson, Executive Director of the HEC Foundation; Marie-Noëlle Semeria, Chief Innovation Officer at TotalEnergies; Céline Brucker, General Manager at L’Oréal France; Joëlle Barral, Director of Research at Google DeepMind; and Julia Peyre, Head of AI Strategy and Innovation at Schneider Electric. Building the future of AI together The scientific perspectives opened by the researchers, professors, and students involved in Hi! PARIS require resources that match their ambition. The trust and engagement of our partners are essential to sustaining this effort. As we look ahead, we remain committed to deepening these relationships and continuing to work hand in hand to address the evolving challenges of artificial intelligence. From sustainability to trustworthy AI, from ethics to real-world applications, we believe the conversation between researchers and industry has never been more essential. The anniversary dinner was a warm and meaningful moment a way to celebrate what we’ve built together and to thank each partner, personally, for believing in our mission. To all our donors: thank you for your support, your time, and your ideas. We look forward to the next chapter and to continuing this journey with you.
The EU Artificial Intelligence Act (AI Act) is the world’s most comprehensive attempt to regulate artificial intelligence, but as Gabriele Mazzini one of its original drafters reminded the audience at the Hi! PARIS Meet Up on the AI Act, it’s also a work in progress. In his talk, Mazzini walked through the logic of the Act, its risk-based foundation, and how recent events have transformed its scope. His reflections offered both a behind-the-scenes view and a forward-looking critique. Why the AI Act focuses on what you do, not what you build The original draft of the AI Act, developed by the European Commission in 2021, was guided by a clear principle: regulate not the technology, but its applications. The idea was to focus on how AI is used, not what it is. As Mazzini explained, this meant identifying risk levels and aligning them with regulatory obligations. Three categories of risk were defined: Prohibited AI applications, such as social scoring or exploitative and manipulative AI. One of the most controversial proposals involved restricting remote biometric identification systems in public spaces. High-risk systems, the heart of the regulation, accounting for around 98% of the legal provisions. These systems would be subject to compliance, certification, and CE marking, similar to how medical devices are regulated. Transparency obligations for systems like chatbots, where the law requires users to be informed when interacting with an AI. According to Mazzini, this is not just a technical issue, it’s about human dignity and respecting the way people relate to machines. Gabriele Mazzini, Architect of the EU AI Act and Research Affiliate at MIT Media Lab | Hi! PARIS Meet Up on the AI Act at VINCI (March 2025) The turning point: ChatGPT and the U.S. influence The final version of the Act preserved the risk-based structure, but it was significantly influenced by two external events: The launch of ChatGPT in October 2022, and the Executive Order on AI from the Biden administration in October 2023, which introduced rules for dual-use foundation models. Together, these developments pushed EU legislators to expand the scope of the AI Act to cover not just applications, but AI tools themselves, especially general-purpose AI models, also known as foundation models. This new chapter introduced two rule sets: Transparency for all models, including documentation requirements and obligations to share information downstream, particularly regarding copyright compliance. Additional obligations for models with systemic risk, including risk assessment, incident reporting, and cybersecurity measures. To determine systemic risk, regulators proposed two criteria: The computational power used in training (mirroring U.S. thresholds like 10^26 FLOPs). And the designation by the AI Office, part of the European Commission, following an approach similar to the Digital Services Act. These rules apply even to open-source foundation models, though some exceptions are allowed. The AI Act isn’t what it started as Since the first draft, the AI Act has changed quite a bit, not just in terms of content, but also in overall complexity. Mazzini pointed out that the number of prohibited use cases has gone from four to eight, with new ones like emotion recognition and categorization, which he described as “vague” and “too broad.” The list of high-risk applications hasn’t exploded, but it’s expanded enough to make compliance more demanding. One big shift is that the regulation doesn’t just focus on applications anymore, it now also covers the AI models and tools themselves. When it comes to general-purpose AI, a lot of the specifics are still being worked out through voluntary codes of practice. That’s led to some debate, especially after a recent letter from EU lawmakers raised concerns about whether those codes are enough to keep up with fast-evolving risks. Meanwhile, governance structures have gotten more complex, both at the EU level and within member states, partly because of the broader scope that now includes general-purpose models. Another important point is the new set of obligations for companies. What used to be called “users” are now “deployers,” and they have more responsibilities, like doing fundamental rights impact assessments and sharing more information. Lastly, Mazzini mentioned that the overlap between the AI Act and other EU laws still isn’t totally clear, and how these different legal frameworks will work together is still being figured out. Hi! PARIS Meet Up on the AI Act at VINCI (March 2025) Less would have been more: Mazzini’s assessment In closing, Mazzini offered a candid reflection: “Less would have been more.” He acknowledged the ambition of the AI Act but emphasized the importance of legal clarity, both for operators who need to comply and regulators who must enforce it. What should come next? Clarity and legal certainty, businesses must understand what they’re required to do, and enforcement must be consistent across EU member states. Sensible interpretations, regulators and courts should aim for realistic, state-of-the-art, and innovation-friendly readings of the law. Harmonized standards, especially for SMEs that lack resources to develop compliance mechanisms on their own. Use-case-based advocacy, companies should engage proactively, using their real-world cases to shape practical interpretations. Impact monitoring, we need data on how the AI Act is working. Is it increasing trust? Creating confusion? Encouraging innovatio or stifling it? “We are dealing with a fast-moving technology,” Mazzini said. “The law matters but so does how we interpret and implement it.” In his view, transparency, pragmatism, and responsiveness will be key to ensuring that the AI Act delivers on its promise without hindering Europe’s AI ecosystem. https://www.youtube.com/watch?v=-jeiKEp-_sM
As part of the Hi! PARIS Cluster 2030 initiative, the Hi! PACE (Programs, Academia, and Course Expansion) positions for the 2025/26 academic year are now open. These positions are the result of a coordinated effort across partner institutions to identify and prioritize teaching and program development needs in artificial intelligence and data science. The positions are aimed at strengthening our academic offering, supporting new program launches, and responding to the growing demand for excellence in AI education. What’s next? Partner institutions can now proceed with recruitment for the listed positions. The roles cover a range of teaching and academic support needs, and are aligned with Hi! PARIS’s strategic vision for inclusive, interdisciplinary, and high-impact education in AI. Open Positions: Artificial Intelligence – ECC (4-year contract) – Université de Téchnologie de TroyesApply here Data Science – ECC (4-year contract) – Université de Téchnologie de TroyesApply here Statistical Learning – Assistant Professor (3+2 years) – ENPCApply here Trustworthy & Responsible AI / RL – Assistant Professor (permanent) – École PolytechniqueApply here Multimodal AI – Monge Assistant Professor (tenure track, 3+3 years) – École PolytechniqueApply here AI for Insurance – Assistant Professor (tenure track) – CRESTApply here Stay tuned for further updates as the Hi! PACE initiative continues to grow. For any questions, please contact: contact@hi-paris.fr