Anthropic adds invisible watermarks to all Claude text to comply with EU AI rules

Anthropic has announced that all new versions of its Claude chatbot will now embed invisible watermarks into every piece of text they generate, as well as attach signed provenance metadata to supported image files. The move is part of the company’s commitments under the EU Artificial Intelligence Act, specifically the Code of Practice on transparency for AI‑generated content that came into effect at the start of August.

For models launched on or after August 2, 2026, Claude weaves a statistically detectable pattern directly into the token stream. Humans should not notice any visible marker, but dedicated detectors will be able to flag whether text likely originated from Claude, even after copy‑and‑paste or light editing. The change applies globally across Anthropic’s products and partner platforms, including API access via AWS, Google Cloud, and Microsoft’s foundry services, rather than being limited to EU users.

Developers and creators are sharply divided. Supporters argue that robust provenance is essential to fight deepfakes, impersonation, and large‑scale spam. Critics counter that forcing a hidden statistical constraint on word choice will inevitably degrade quality—especially for code—and that documents lightly edited with Claude could be labeled “AI‑generated” in ways that complicate education, journalism, and professional writing. The underlying technique is based on academic work showing how to bias an LLM’s token sampling so that machines, but not humans, can detect the pattern. Whether rivals like OpenAI and Google adopt similar model‑level watermarks under regulatory pressure will be a key test for the emerging AI transparency regime.