Anthropic's text watermark alters word probabilities to embed a fingerprint, which could degrade Claude's writing, despite its claim of no impact on quality (John Gruber/Daring Fireball)

Anthropic has provided insight into the workings of its AI model Claude's new text watermark, which it implemented in line with Article 50 of the European Union's AI Act. This rule came into force in August, mandating that generative AI systems mark outputs as artificial whenever feasible. The watermark, according to Anthropic, doesn't involve hidden characters or metadata tags, as plain text lacks space for such additions.

Rather, it embeds a subtle statistical pattern in the model's word choices during generation. When multiple natural sentence completions exist, Claude uses a secret mathematical key during its selection process, rather than random choices. The mark, Anthropic claims, doesn't change the meaning, quality, or readability of the response, and can endure through copying, pasting, and certain edits.

This watermark marks Claude models that were launched on or after August 2 globally, not just in the EU. It covers Claude apps, API, Claude Code, Cowork, and Tag, as well as certain deployments through AWS, Google Cloud, and Microsoft Foundry. Anthropic is working to extend the capability to older models, though no specific timeline has been given.

Anthropic also pointed out that a detected watermark indicates Claude may have processed the content, not necessarily authored it. This is because Claude is also used for text translation, summarization, proofreading, or editing. The company also noted that the watermark may not appear on very brief passages or narrow, factual text, as there are fewer opportunities for the pattern to embed itself in.

Anthropic hasn't disclosed the technical implementation of the watermark, and independent researchers have shown that watermark schemes like this can sometimes be spoofed or removed through heavy rewriting or paraphrasing. The requirement is driven by the EU AI Act's aim to make synthetic content distinguishable, with penalties of up to €15 million or 3 percent of a company's global annual turnover for non-compliance.

Anthropic has adhered to Article 50(2) of the EU AI Act as both a model provider and a system provider. Deployers of Claude in their own products are advised to independently evaluate their own Article 50 obligations.