Thomas Souverain

AI Ethicist, Dr. in Philosophy of Computer Science — Paris • Oxford • Stanford

View on GitHub

Programming to Integrate AI Ethics - Personal Works and Team Management

  1. CYBERSECURITY - Making European law on generative AI detection operational (2025-2026) At the French Alternative & Atomic Energy Commission, I am currently developing more secure and transparent watermarking approaches for Large Language Models (LLMs) on one of the most powerful European data centers, Jean Zay. I was recently consulted by the European Commission to write EU guidelines for LLM watermarking, based on the gap with the current EU legislation AI Act. The theoretical part is pre-printed in the FaCCT paper. Currently, I lead a team of two data-scientists to develop specific Red Teaming tests for French LLMs in high schools.

  2. EXPLAINABILITY - Stanford Computer Science Department (2023) During the Winter Semester 2023 (01-04/2023), I visited Stanford and tested a specific neural network architecture (graphical). The purpose was to integrate a causal interventionist logics, to make financial agents able to constraint with their domain knowledge on causal ancestors (e.g. the gender influence on the type and level of education).
    GitHub Repository

  3. FAIRNESS - Detecting and Mitigating Discriminations in Loan Lending (2022) I built a package to make algorithmic fairness choices more accessible to lay users in binary classification (e.g. scoring, loan lending), which went to final of the 2022 Hackathon of the Monetary Authority of Singapore. My method reweighing disadvantaged clients is described in the Topoi paper, and accessible in this GitHub Repository.

Trainings on AI Ethics (2021-2026)

I developed trainings for current and future AI practicioners, being either governance leaders, computer scientists, engineers, lawyers. Towards French Universities and engineering schools, I built teaching for masters’ and continuing education. I developed full trainings combining applied ethics, European regulatory frameworks, and hands‑on ML exercises. See for instance a practical coding exercise I created on algorithmic loan lending on Kaggle, to make data scientists and jurists work together and become aware of fairness choices behind explainable AI tools.

Guidance of Product Teams - Embedding Ethical Practices

  1. For the French Ministery of Education (2025-2026), I lead experiments comparing higher and secondary education to measure the impact of generative AI on studying, learning, and criticizing. I also built specific contents for teachers and students, raising awareness on the real techniques and ethical stakes behind AI.

  2. At the French Employment Agency (France Travail, 2023-2024), my quantitative and qualitative experiments helped redisign the platform of AI discrimination alerts, to make it more efficient and understandable for end-users

  3. In the fintech start-up DreamQuark (2019-2023), I advised the CEO with an executive strategy to integrate explainability and fairness pipeline in AI loan lending, churn detection, and product recommendation.

Executive Stategies – White Papers to guide AI companies’ leaders and operating officers

  1. Souverain, T., Grinbaum, A., & Klein, E. (2026). Watermarking LLMS: feasibility, constraints, and strategies.
    CEA Paris-Saclay. Download

  2. Souverain, T. et al. (2023; re-ed. 2025, French). Integrating Ethics in AI Systems.
    Hub France IA.
    Download

  3. Souverain, T., & Meric, N. (2021). State of Ethical AI: Challenges posed by ethics to companies developing AI.
    Report, 110 p. Analysis of stakes, international strategies, and applications to integrate Ethics into AI development. Download