What Anthropic has said about watermarking

Article 50(2) of the EU AI Act requires providers of covered systems to mark generated or manipulated audio, image, video and text output in a machine-readable form, subject to stated exceptions and technical feasibility. The obligations apply from 2 August 2026. Providers of covered content-generating systems already on the market before that date have until 2 December 2026.

The law does not prescribe a watermarking method. Anthropic has signed the voluntary Code of Practice and uses the same two dates in its rollout wording. Its decision to apply marking worldwide is a product choice described on the help page, not a worldwide extension of EU law.

Anthropic defines support by launch date and does not name one marking-enabled model. The four model assignments in our byte-level scan all predate 2 August. All four therefore fall into the older-model group, where Anthropic says support is still in progress.

This does not prove that those models are unmarked. Anthropic may retrofit an older model and update the page later. It means that a claim about a specific current model needs evidence that is not in the published documentation.

We also could not verify the model names independently. The collection harness assigned them and each subagent repeated them, but we could not inspect an API response header.

In June 2026, a researcher found that the Claude Code client altered the date inside its system prompt under certain base-URL and timezone conditions. The client changed an apostrophe to one of three lookalike characters and could change hyphens in the date to slashes.

That finding concerns input sent to the model. The new help page concerns output generated by the model. It does not tell us how the output watermark works.

Our scan found none of the tested hidden-Unicode classes in the 457,045 characters that reached disk. That result covers those 120 files and that collection path only. It does not rule out a mark carried by word or token choice, and it does not cover six requested non-ASCII space characters that reached disk only as ordinary spaces.

Research systems such as the Kirchenbauer green-list method and SynthID-Text alter token selection during generation. Their marks can be present in plain ASCII text, so a character scan would never see them. They show that a text watermark need not use unusual characters.

They do not describe Claude. Anthropic names no mechanism and publishes no detector. A Unicode cleaner can remove unusual characters, while a rewrite can change word choices, but neither operation can certify that an official Claude check would now pass. There is no published official check to compare against.

Anthropic's technical documentation and detector would answer this directly. Until those arrive, an independent test would need fixed prompts, models on both sides of the launch-date rule, raw responses and a method another team can rerun. We have collected the earlier-model side: four assignments on one route. We found no hidden Unicode there. We did not discover or remove Claude's watermark.

Read the byte-level scan, or read how this site's published watermark works.