TL;DR
- Planned marking: Anthropic says future Claude models will mark generated text and supported files so readers and organizations can check whether Claude helped create or process them.
- Text signal: During generation, Claude can favor some equally suitable next words using a key, leaving a statistical pattern that a detector can use to estimate Claude involvement.
- File record: Supported image files can carry a signed C2PA Content Credential, a metadata record that says Claude made or processed the file.
- Limits: A detected mark indicates possible Claude involvement, not authorship or truth; short or factually fixed passages such as code, and text changed by heavy editing or rewriting, give the detector less evidence.
- Timing: Starting Aug. 2, 2026, the EU AI Act requires providers to mark specified synthetic text, audio, images, and video so they can be detected as AI-made where technically feasible; Anthropic’s older-model rollout and public detector remain future work.
Anthropic is planning two ways for users and organizations to check whether Claude handled an output: a statistical pattern in generated text and signed C2PA metadata that records Claude’s involvement in other created files. The text mark would be read from the words, while the file credential would be read from the file’s metadata.
What the Text Mark Changes
Claude generates a response one token at a time, choosing among words that fit the text so far. When several choices would work equally well, Anthropic says its watermark will use a private key and the preceding words to influence the source of that randomness. The selected words should still read normally, but their sequence can carry a statistical pattern that a detector checks with the key. Anthropic says the method adds no visible characters and no extra tokens.
That makes the text watermark narrower than a general authorship checker: The underlying test estimates whether Claude was involved in at least part of the passage using a Claude-specific key and the statistical pattern in generated words. It is a provider-specific signal, not a universal detector that will work for every AI system.
The signal also depends on how many meaningful choices the model had. Short passages, factual sentences with one correct answer, and code often give the mark little room to register. Claude’s own examples say proofreading can leave too few changed words to identify Claude reliably, while a translation generated by Claude gives the model many more choices to mark.
Not an Authorship Certificate
Anthropic says a detected mark can show that Claude was probably involved, but it cannot distinguish between Claude writing a passage and Claude heavily editing one. It also carries no identifying information about the user, organization, or chat, and does not change ownership or legal responsibility under Anthropic’s terms.
The reverse inference is also unsafe. A passage without a detectable mark is not thereby definitively considered human-written. A full rewrite, paraphrase, or other model can replace the choices that carried the original signal. The SynthID-Text study that Anthropic cites likewise finds that detection depends on text length and the amount of choice available during generation, and describes paraphrasing and signal-scrubbing as vulnerabilities.
The study is useful evidence about the technique. Google researchers evaluated SynthID-Text with Gemini, Gemma, and Mistral models, including a live experiment involving about 20 million Gemini responses. They reported no statistically significant quality difference between watermarked and unwatermarked responses in that experiment. Those results show that generation-time watermarking can be deployed at scale, but they do not establish Claude’s future detection accuracy, false-positive rate, or resistance to editing.
Text Marks and File Credentials Are Different
Text watermarking lives in the choices made while a model generates words. File provenance lives in the file itself. For supported outputs such as PNG, JPG, and SVG files, Anthropic says Claude will attach a small, cryptographically signed C2PA Content Credential to the metadata. A C2PA-aware tool can read that record and report that Claude made or processed the file.
The C2PA standard treats a Content Credential as a signed record of a file’s origin and subsequent actions, not as a verdict that the file is true or that a person authored it. It can also be incomplete or disappear when a file is stripped of metadata or moved through an unsupported workflow. That is a different failure mode from a text mark being weakened by short passages or rewritten words.
The distinction matters for organizations that receive both prose and generated files. A text check asks whether Claude’s generation pattern is present in the words. A file check asks whether a signed provenance record survives in the file. Neither check is a universal detector for every AI system, and neither can recover a complete history after unsupported handling removes the signal.
As of Aug. 17, 2026, Anthropic’s support guidance says models launched in the European Union on or after Aug. 2, 2026, are to carry marking at launch. It says the company plans worldwide marking because it cannot yet scope the feature by region, while older models are being updated over the following months. The same guidance makes some product, cloud, file, and builder coverage conditional.
Anthropic’s public materials do not name a live marked Claude model or output or provide a details about a public detector endpoint. The stated Aug. 2 rule so far is a launch commitment, not a verified list of currently marked models. Users therefore have no documented public detector to query yet.
Why Anthropic Is Timed to August 2
The timing follows the European Union’s AI Act. Article 50 requires providers of systems that generate synthetic text, audio, images, or video to mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated, as far as technically feasible. It separately requires deployers to disclose certain AI-generated text published to inform the public on matters of public interest, with an exception for content that has undergone human review or editorial control and has a responsible publisher.
The European Commission’s explanation of the marking code draws an important line: Article 50’s transparency requirements are legal obligations, while adherence to the Code of Practice is voluntary. The code gives signatories a recognized way to demonstrate compliance, but signing it is not the source of the legal duty.
Anthropic’s choice to apply the plan globally is therefore a provider implementation decision made against a European legal deadline.
Until Anthropic documents a public detector and live model coverage, a Claude mark will be a provider-specific clue and C2PA credentials a file-history record that can be removed from metadata.