The Short Answer

As of August 2, 2026, Claude embeds an invisible watermark in every piece of text it generates and attaches signed provenance metadata to supported file types. The policy is global, applies across every Claude product and API surface, and cannot be turned off by users. It was triggered by the EU AI Act's Article 50 transparency requirements, but Anthropic chose to apply it worldwide rather than limit it to Europe.

Why Anthropic Did This Now

The immediate regulatory trigger is EU AI Act Article 50, which became enforceable August 2, 2026 for newly launched AI systems. Anthropic has signed the European Union's Code of Practice covering AI-generated content. Companies that comply with that code of practice receive a presumption of regulatory conformity, which shifts the legal burden away from Anthropic in enforcement proceedings. The financial stakes are not trivial: companies that violate the AI Act face penalties of up to €35 million or 7 percent of their worldwide yearly turnover, whichever is higher.

Anthropically could have confined the watermark to EU users. It did not. Anthropic says those markings will apply worldwide, not only to users in Europe.

How the Watermark Actually Works

The system uses two distinct mechanisms depending on the type of output.

For text: The text watermark works by statistically biasing Claude's word choices according to a key held by Anthropic. Individual choices look unremarkable, but across enough text, the pattern becomes detectable. You won't see it, and it doesn't change the meaning, quality, or readability of Claude's response. Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing.

For files: A help center article confirms that Claude models launched on or after August 2, 2026 attach signed C2PA provenance metadata to supported file types like .svg, .png, and .jpg. That signature can be checked to verify whether the file has been tampered with since Claude generated it. It can be stripped through re-saving or format conversion, which is why the two systems are meant to complement each other.

Where it applies: Claude markings cover output from supported models everywhere you use Claude, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. This includes third-party products built on the Claude API.

What the Watermark Cannot Do

Anthropically is candid about the limits, and those limits matter.

The existence of a watermark or metadata "is not fully conclusive" that the content was completely generated by AI, only that it was processed by Claude models. That is a significant caveat. Text originally written by a human but entered into Claude for proofreading, translation, or summarizing would carry a watermark after it is processed.

The reverse is also true. There will be no detectable signal on content generated from older models, very short passages, heavily paraphrased text, or files whose metadata was stripped through screenshots or format changes. An engineer at Anthropic acknowledged the obvious limitation: "it's not perfect, you can edit it, but it's a first step."

Paraphrase defeats the signal, which means the watermark is better described as a transparency signal than a detection guarantee.

Detection Tools: Coming, Not Here Yet

Anthropic has announced "a text detection API that you can use yourself," but as of this writing the tool has not shipped. No pricing, rate limits, or access tier have been given. Until that API is live, third parties and end users have no official way to check whether a given piece of text carries a Claude watermark. Anthropic has committed to publishing more technical documentation alongside the detection tool.

The Comparison Landscape

Claude is not operating in a vacuum. Google has been watermarking its AI-generated images since 2023 and has extended that to text, audio, and video. OpenAI has reportedly had the capacity to watermark ChatGPT text for years, but has not done so at scale. Claude is the first major frontier AI lab to deploy production-scale text watermarking across all its products at once.

For developers using Claude Code or building on the Claude API, note that the watermark applies to code output as well. "All Claude generated text will have this embedded watermarking," which includes code snippets and generated documentation, not just prose.

Who Should Pay Attention

Developers and API users: If you deploy Claude in your product, you must "independently assess what Article 50 requires" for your own compliance obligations. The watermark is Anthropic's signal, not a substitute for your own disclosure requirements.

Writers and knowledge workers: Lawyers, academics, researchers, and writers who use Claude to copy, edit, or revise work that is substantially their own now face the prospect of that work carrying an AI marker they did not choose. A file that a person wrote and lightly polished with Claude could return a positive detection when run through Anthropic's forthcoming tool or another third-party system. This is not a theoretical concern.

Students: Students and professionals trying to pass off AI-generated work as their own may soon have a harder time getting away with it, once the detection API ships. Whether institutions will actually use Anthropic's detection tool, rather than existing third-party services, remains an open question.

Teams using AI writing tools: If your workflow involves using Claude alongside tools like Grammarly, QuillBot, or Wordtune for post-processing, be aware that the watermark may persist through light edits but degrade through heavy paraphrasing. Running Claude output through a paraphrasing tool would likely strip or damage the signal, though Anthropic has not published specifics on the threshold.

The Opt-Out Question

There is no opt-out. This global policy, without an opt-out, is largely driven by the EU AI Act's Article 50, which mandates transparency for AI outputs. Watermarking is applied at the model level, which means it will be present no matter which Claude product or surface the text comes from. Anthropic has not indicated it will offer enterprise opt-outs even for use cases where watermarking creates operational complications.

Practical Takeaways

  • The watermark is invisible, does not affect output quality, and travels with copy-pasted text.
  • A watermark means Claude touched the content. It does not mean a human did not author it.
  • No watermark does not mean the content is human-written. Older models, short outputs, and paraphrased text will not carry a signal.
  • A detection API is coming. No timeline or pricing has been confirmed.
  • If you use Claude to edit or revise human-written material, that output will carry a mark. Factor this into your disclosure policies.
  • Developers building on the Claude API need to review their own Article 50 obligations independently of Anthropic's compliance.

The watermark system is a genuine first step toward machine-readable AI transparency, with real gaps that Anthropic itself acknowledges. Treat detection results as a signal worth investigating, not a verdict.