AI watermarks in Europe do not yet meet all four legal tests

A review of watermarking methods finds no current approach that is simultaneously reliable, interoperable, effective and robust under the European Union’s AI Act.

Editorial illustration — not from the study.

The paper develops a taxonomy of watermarking methods based on when they are applied in the life cycle of a large language model, including before or during training and at the stages of token selection and sampling. It also connects the AI Act’s requirements with existing evaluations of watermark detectability, robustness and language-model quality.

Because interoperability has received little theoretical treatment in this research area, the authors propose three normative dimensions for assessing it. Comparing current methods with the four European criteria, they conclude that no approach yet satisfies all of them and recommend more research on watermarks embedded in the underlying architecture of language models.

What current watermarks achieve

The analysis finds that current large-language-model watermarking methods do not meet all four standards named in Article 50 and Recital 133 of the European Union’s AI Act: reliability, interoperability, effectiveness and robustness. The authors classify watermarking methods by where they are introduced in the model’s life cycle and distinguish evaluations of the watermark itself from evaluations of the model’s output quality. They also propose three dimensions for thinking about interoperability, which they say remains largely underdeveloped in existing research.

Why the legal gap matters

The findings show that the legal requirement to mark and detect AI-generated output does not translate neatly into a single established technical solution. The comparison provides a framework for assessing whether watermarking methods can meet the European Union’s stated criteria, while highlighting interoperability as an unresolved part of that assessment. The paper’s recommendation to study watermarks built into the low-level architecture of language models points to one direction for future research, rather than a method already shown to meet the law’s requirements.

Evidence and limits

This is an analytical article that proposes a taxonomy, interprets the AI Act’s criteria, and compares current watermarking methods using existing evaluations of detectability, robustness and model quality. The abstract does not report a new experiment, a sample size or numerical performance results. Its conclusions are limited by the rapidly changing and varied watermarking field, and by the fact that interoperability has not yet been consistently defined or evaluated in this area.

High-mounted cameras spot pedestrians on the device at busy intersections

Researchers introduced a 13,314-image dataset for testing AI systems that view urban intersections from above. The results show that detection can run on edge devices such as the Jetson Orin Nano, although pedestrians remain harder to detect than vehicles and performance changes with scene conditions.

5 hours ago·3 views

AI & Computing

Where smaller AI models can work alongside—and compete with—larger ones

A survey examines the roles smaller AI models can play alongside large language models. It focuses on both collaboration and competition, including how smaller systems may help researchers and businesses use computing resources more efficiently.

22 hours ago·8 views

AI & Computing

An AI attack can fool one client's system while sparing another

Researchers describe an attack on personalized federated-learning systems that aims to fool selected client models without disrupting designated “friend” models. In the main test setup, it succeeded against at least 88% of target cases while preserving at least 81% accuracy on friend models, though performance varied with the information available to the attacker.

6 days ago·56 views