Anthropic deploys invisible text watermarking to meet EU AI Act mandates
The AI lab is implementing a statistical signature in Claude models to identify machine-generated content as transparency regulations take effect.
Anthropic has introduced a text watermarking system for its Claude AI models to comply with the European Union's AI Act. The move ensures the company meets new transparency requirements designed to help regulators and users distinguish machine-generated text from human writing.
The system works by subtly influencing the model's choice of tokens during the sampling process, creating a statistical signature that can be identified using a digital key. According to Anthropic, the technique is based on a method previously introduced in Google DeepMind's SynthID-Text paper. The company states that the watermark is designed to be imperceptible to human readers and has no statistically significant impact on the performance or cost of the models.
Regulatory pressure and timelines
The implementation is a direct response to the EU AI Act's Transparency Code, which imposed new obligations on AI providers starting August 2, 2026. To meet these deadlines, Anthropic has integrated the technology into all Claude models released on or after that date. For older models already in circulation, the company is managing a transition period to upgrade the systems, with a final deadline of December 2, 2026.
Unlike image or video watermarking, which can rely on fixed pixel grids, text requires a probabilistic approach. Anthropic's method focuses on "inconsequential" word choices to embed the signal. However, the company noted that this application is not uniform across all types of content. "Watermarking is sparser on factual passages where there are fewer choices that can be made without decreasing the accuracy of the text," Anthropic stated.
The detection arms race
This shift toward invisible regulatory compliance highlights the growing struggle to police "AI slop" and misinformation. While the watermark provides a layer of accountability, it is not a permanent seal. Anthropic admitted that while light editing is unlikely to remove the signature entirely, "a complete rewrite where every word is replaced will."
This limitation underscores an ongoing arms race between AI detection tools and evasion techniques. As regulators demand more robust provenance for digital content, the reliance on statistical signatures suggests that detection remains a game of probability rather than absolute certainty. For now, the industry is moving toward a standard where AI labs take proactive responsibility for marking their output, even if those marks can be erased by a determined user.