Anthropic says it will watermark Claude's output. Nothing is marked yet.

A help page published without announcement sets out how Anthropic will comply with the EU's new marking rule. Anthropic's pledge binds models launched in the EU on or after August 2, and Claude has not launched one since July 24.

Agent-drafted, human-editedPublished August 11, 2026

Reviewed by Trevor Brown

Anthropic has published how it intends to mark Claude's output as AI-generated. For now, the policy marks nothing.

A help center page carrying an August 10 modification date, posted without an announcement, says Claude models "launched in the EU on or after August 2, 2026 will support machine-readable marking at launch."1 Anthropic's most recent model launch was Claude Opus 5 on July 24, on both its newsroom and its platform release notes, checked Monday.2 Every model in production today therefore falls under the page's transition clause rather than its commitment.

Why it matters: the EU's requirement that AI-generated content be marked in machine-readable form took effect August 2. Until detection exists, no one outside Anthropic can check whether the watermark works.

How it would work: two techniques. The page says Claude "weaves an imperceptible watermark directly into the text itself," applied at the model level so it travels with copied text and, in Anthropic's words, "may persist through some editing." Generated files of supported types, such as .svg, .png and .jpg, get signed provenance metadata following the C2PA open standard.1

Anthropic says the marks would cover the API, Claude, Claude Code, Claude Cowork and Claude Tag, and apply "wherever Claude is offered, worldwide" rather than only in the EU. Watermarks would reach models served through AWS, Google Cloud and Microsoft Foundry, though the page cautions that the file metadata "may not be supported on every platform."1

The candid part: the page spends more words on limits than on mechanics. "A detected mark provides a signal that content was processed by Claude, but is not fully conclusive," it says, noting that Claude is often used to rework material that originated elsewhere. Absence proves as little: a mark may be missing because the model predates marking, because text was heavily edited or too short for a reliable signal, or because metadata was stripped by a screenshot or a format conversion.1

Detection, the thing that would make any of this checkable by outsiders, does not exist yet. Anthropic says it will "share details on detection mechanisms in forthcoming technical documentation," and gives no date.1

A rival ships public detection. Anthropic ships none. OpenAI runs a verification tool that checks uploaded images and audio for C2PA Content Credentials and SynthID watermarks and reports whether either is found. It does not cover text, the format Anthropic's watermark targets, so neither company can yet prove a piece of writing came from its model.

Provenance & disclosure

AI use

Agent-drafted, human-edited

Disclosure

Drafted by Claude from Anthropic's August 10 help center page, Anthropic's newsroom and Claude Platform release notes checked the same day, the Official Journal text of Regulation (EU) 2026/1744, the European Commission's Code of Practice materials, the text of AI Act Article 50, and OpenAI's public provenance verification tool, then edited by Trevor Brown. Every claim about how the marking performs is Anthropic's own and cannot be independently tested until detection tooling exists. Conflict of interest, stated plainly: Still in the Loop's drafts are written by Claude, so this story reports on a policy that would mark this publication's own draft text. It was drafted by Claude Fable 5, a model made by the company it covers.

Reviewed by

Trevor Brown — Editor-in-chief

Sources

1.Official sourceThe signing of the Code of Practice, the "launched in the EU on or after August 2, 2026 will support machine-readable marking at launch" commitment, the transition clause for earlier models, the two marking techniques and the "imperceptible watermark" language, the .svg / .png / .jpg file types and the C2PA standard, the product and cloud-partner coverage, the "wherever Claude is offered, worldwide" scope, the "not fully conclusive" limitation and the reasons a mark may be absent, and the deferral of detection details to "forthcoming technical documentation," from Anthropic's page, How Claude marks AI-generated content, modified August 10.
2.Official source