Watermarking Large Language Models in Europe: Interpreting the AI Act in Light of Technology

To foster trustworthy Artificial Intelligence (AI) within the European Union, the AI Act requires providers to mark and detect the outputs of their general-purpose models. The Article 50 and Recital 133 call for marking methods that are ''sufficiently reliable, interoperable, effective and robust''. Yet, the rapidly evolving and heterogeneous landscape of watermarks for Large Language Models (LLMs) makes it difficult to determine how these four standards can be translated into concrete and measurable evaluations. Our paper addresses this challenge, anchoring the normativity of European requirements in the multiplicity of watermarking techniques. Introducing clear and distinct concepts on LLM watermarking, our contribution is threefold. (1) Watermarking Categorisation: We propose an accessible taxonomy of watermarking methods according to the stage of the LLM lifecycle at which they are applied - before, during, or after training, and during next-token distribution or sampling. (2) Watermarking Evaluation: We interpret the EU AI Act's requirements by mapping each criterion with state-of-the-art evaluations on robustness and detectability of the watermark, and of quality of the LLM. Since interoperability remains largely untheorised in LLM watermarking research, we propose three normative dimensions to frame its assessment. (3) Watermarking Comparison: We compare current watermarking methods for LLMs against the operationalised European criteria and show that no approach yet satisfies all four standards. Encouraged by emerging empirical tests, we recommend further research into watermarking directly embedded within the low-level architecture of LLMs.

Implementing Fairness: the view from a FairDream

In this paper, we propose an experimental investigation of the problem of AI fairness in classification. We train an AI model and develop our own fairness package FairDream to detect inequalities and then to correct for them, using income prediction as a case study. Our experiments show that it is a property of FairDream to fulfill fairness objectives which are conditional on the ground truth (Equalized Odds), even when the algorithm is set the task of equalizing positives across groups (Demographic Parity). While this may be seen as an anomaly, we explain this property by comparing our approach with a closely related fairness method (GridSearch), which can enforce Demographic Parity at the expense of Equalized Odds. We grant that a fairness metric conditioned on true labels does not give a sufficient criterion to reach fairness, but we argue that it gives us at least a necessary condition to implement Demographic Parity cautiously. We also explain why neither Equal Calibration nor Equal Precision stand as relevant fairness criteria in classification. Addressing their limitations to warn the decision-maker for any disadvantaging rate, Equalized Odds avoids the peril of strict conservatism, while keeping away the utopia of a whole redistribution of resources through algorithms.