Claude now marks text and files processed by its models

On August 11, 2026, TechCrunch reported Anthropic's new system for marking content processed by Claude. Anthropic says the system applies worldwide across Claude products. It puts an imperceptible watermark into text at the model level and adds signed provenance metadata to supported files. The central distinction is important: the mark can show that Claude processed content, but it does not prove that Claude wrote it.

What Claude marks

Anthropic describes two mechanisms on its official help page. For text, Claude inserts an imperceptible watermark directly into the text. Marking happens at the model level. Anthropic says this means the watermark is present regardless of which Claude product or surface produced the text.

The text mark travels when someone copies and pastes the text elsewhere. Anthropic says it may also persist through some editing. The company does not describe the exact technique. Technical documentation about detection is forthcoming, so there is no basis yet for describing a particular algorithm.

Files use a different mechanism. Claude adds signed provenance metadata that follows the C2PA open standard. C2PA stands for the Coalition for Content Provenance and Authenticity. Anthropic specifically names .svg, .png, and .jpg, as well as other formats where supported. This metadata records provenance in the file rather than placing the text watermark into its visible pixels.

Coverage depends on model age. New models launching August 2, 2026 or later support marking at launch. Support for earlier models is still in progress. Anthropic says the policy applies worldwide across Claude products. Its page does not state whether users can opt out.

The timing also has a regulatory context. TechCrunch linked the change to the EU AI Act's Transparency Code, effective August 2, 2026. The code requires AI companies to mark AI-generated or AI-edited content in a form other systems can identify.

What this changes when choosing a writing tool

People evaluating Claude should treat marking as a record of model involvement, not an authorship verdict. That difference matters for ordinary editing. If you give Claude a paragraph you wrote and ask it only to proofread or translate it, the resulting document is marked. Detection therefore cannot, by itself, establish who composed the original argument or sentences. Anthropic says a detected mark is a signal that content was processed by Claude and is not fully conclusive.

The same distinction is useful when comparing workflows across the AI Writing category. A writer might generate a draft in ChatGPT, check language in Grammarly, paraphrase material with QuillBot, or translate it through DeepL. Anthropic's announcement does not make claims about how those products mark content. It does show why review policies should separate generation, proofreading, paraphrasing, and translation rather than treating every detected Claude mark as proof of original authorship.

For a technical buyer, the practical question is what a downstream system can infer. The mark can support a claim that Claude handled the content. It cannot reliably answer what Claude did, how extensive the change was, or which parts began as human writing. Any moderation, education, publishing, or workplace process built around the signal needs to preserve that processing-versus-authorship gap.

What is available today, and what is not

Marking is available at launch for Claude models released on or after August 2, 2026. Earlier-model coverage is not complete; Anthropic describes it as in progress. Text watermarking works across Claude products at the model level. C2PA provenance applies to supported file formats, including the formats Anthropic names.

Detection tooling for users and third parties has not shipped yet. Anthropic says it will support detection and will publish technical documentation, but both remain forthcoming. That means the marks can be applied now without a generally described detection workflow from Anthropic being available today.

The signals also have stated technical limits. Heavily edited or paraphrased text can lose detectability. Short passages may not contain enough signal. Removing file metadata removes the C2PA mark, while unsupported platforms and formats carry no mark. These are limits on reliability and coverage, not evidence about how a person created a document. Anthropic provides no accuracy rate or false-positive rate on the page.

The next items to watch are Anthropic's detection tools, its technical documentation, and completion of support for earlier models. Until those ship, the sound interpretation remains narrow: a detected mark signals processing by Claude, not proof that Claude authored the content.