How Claude marks AI‑generated content

Four promises define how Anthropic will put the Code of Practice into practice. They turn one legal obligation into a concrete, machine-readable reality — and they hinge on a single date.

1. i

   Marking from day one

   New Claude models launched in the EU on or after August 2, 2026 support machine-readable marking at launch. Generated text carries an embedded watermark; generated files include digitally signed provenance metadata where supported.

2. ii

   Everywhere you use Claude

   Marks apply to output from supported models across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag — wherever Claude is offered, worldwide. Some platforms or features may not support certain marking types.

3. iii

   Detection, supported

   Anthropic will support users and other third parties in detecting Claude’s marks, as the Code requires — and will share the details in forthcoming documentation.

4. iv

   Earlier models in transition

   The law includes a transition period for Anthropic models launched before August 2, 2026 — and work to add marking support to those models is under way.

Two complementary techniques do the work: an imperceptible watermark woven into text at the model level, and a signed provenance label attached to generated files. Try both.

Technique 1 · Text

The watermark you can’t see

When a supported Claude model generates text, it weaves an imperceptible watermark directly into the words. You won’t see it. It doesn’t change the meaning, the quality, or the readability of the response — and because it is applied at the model level, it is present no matter which Claude product or surface the text comes from. The visualization below reveals what only a detector can perceive.

Claude · generated response Plain to every reader

The golden hour arrives twice a day — once just after sunrise, and again before sunset — when the sun hangs low and its light travels through more of the atmosphere. The air scatters the shorter blue wavelengths and lets the longer reds and ambers through, so everything the light touches seems warmer, softer, and briefly gilded. Photographers plan their days around it.

wm::claude-text-v1 · pattern confidence 99.7% · origin: model-level

Reveal the watermark

Then simulate what happens to the signal:

Copied & pasted Lightly edited Heavily paraphrased

Watermark signal

Because the watermark is part of the text, it travels with the words when they’re copied and pasted elsewhere — and may persist through some editing.

A visualization. The real watermark is imperceptible: the text above looks, reads, and means exactly the same whether it is marked or not.

Technique 2 · Files

The signed label on every file

When Claude generates a supported file type — such as .svg, .png, or .jpg — it attaches signed provenance metadata following the C2PA (Coalition for Content Provenance and Authenticity) open standard, used across the industry to record where content comes from. A signed label signals that a file was processed by Claude — and lets you detect whether the file has been tampered with. Inspect one.

.svg .png .jpg

dawn-study.svg · 1.8 KB · generated by Claude

Inspect provenance Simulate tampering

An illustrative manifest. The structure mirrors the C2PA open standard; detection tooling arrives with Anthropic’s forthcoming technical documentation.