EU AI Act reportedly requires detectable AI text watermarks but removal remains trivial

The EU AI Act’s Article 50 introduces a legal requirement for AI-generated text to be watermarked, enforceable from August 2026. This targets providers like Google, Anthropic, and OpenAI, which must ensure their outputs are detectable as artificially generated. The mandate reflects broader concerns about AI-generated content’s authenticity, but the technical challenges of text watermarking complicate compliance. Unlike images, where noise can hide watermarks, text’s compressed nature leaves little room for imperceptible modifications. Any watermarking scheme must balance detectability with output quality, a trade-off that current methods struggle to resolve.

Google’s SynthID exemplifies one approach, using mathematical relationships between tokens to embed a watermark. The method assigns scores to tokens based on their context and selects tokens with the highest scores during generation. While this is computationally efficient for detection, it is also fragile. Minor edits, such as rephrasing or synonym substitution, can disrupt the token patterns, rendering the watermark undetectable. The reliance on subtle statistical relationships means that even high-quality human edits may break the watermark, limiting its practical utility for enforcement. This fragility is inherent to text watermarking, as the medium lacks the redundancy of images or audio.

Alternative methods, such as Unicode trickery or rule-based constraints, face similar limitations. For example, enforcing a pattern like 'every fifth letter is an e' would degrade output quality, while letting the model adapt the watermark consumes computational resources. The EU’s requirement for free watermarking services further complicates matters, as providers must scale detection without prohibitive costs. The tension between regulatory goals and technical constraints suggests that watermarking may serve as a deterrent rather than a foolproof solution. Engineers must prepare for compliance while acknowledging that watermarks are unlikely to provide robust protection against misuse.

The broader implication is that watermarking alone may not satisfy Article 50’s requirements. Providers might need to combine watermarks with other detection methods, such as model inference checks, but these are computationally expensive and prone to false positives. The EU’s enforcement timeline leaves little room for refining these techniques, forcing providers to adopt imperfect solutions. For engineers, this means designing systems that can adapt to evolving regulatory and technical landscapes, while accepting that watermarking is a temporary and easily bypassed measure.