The newsroom flagged a Claude Help Center article as possible evidence that Anthropic now watermarks all Claude output. That would be high-value safety and governance news: EU AI Act Article 50 transparency rules became applicable on 2 August 2026, and a provider-level mark would travel through every product that wraps Claude, including the API, Claude Code, and cloud marketplaces.
The primary record confirms a real rollout, not a rumor. Anthropic signed the Commission’s Code of Practice on Transparency of AI-Generated Content as a Section 1 provider, published a 14 August 2026 technical explanation, and later documented watermarks in Claude Platform release notes. The overclaim is “all output.” Marks apply to supported models and supported file types, they are machine-readable rather than visible badges, and they are probabilistic signals of Claude involvement rather than certificates of authorship.
What we checked
- Does Claude now watermark all generated outputs, which surfaces are covered, and when did this policy become active? partly confirmed. Supported Claude models launched in the EU on or after 2 August 2026 mark generated text at launch, worldwide, across Claude Platform (API), Claude, Claude Code, Claude Cowork, Claude Tag, and cloud partners. Older models are still being retrofitted, C2PA covers only supported generated media files, and Anthropic itself lists cases where a mark will be missing or weak. [1][4][5]
- Is this a new change or just an updated support article explaining existing behavior? partly confirmed. It is a new 2026 compliance rollout, not a restatement of old Claude behavior. EU Article 50 obligations applied from 2 August 2026; Anthropic signed the Code of Practice in July 2026, published the technical note on 14 August 2026, and documented watermarks in the 1 September 2026 Fable 5.1 release notes. The Help Center article is the living policy page and was showing as updated the day before this review. [2][4][5][1]
- Is Claude text-watermark detection publicly available to any user or API customer? refuted. Text detection is in private preview for eligible organizations required under EU law, plus enterprises that must verify watermarking for their own Article 50 compliance. The public Claude Content Checker only reads C2PA credentials on files and explicitly does not check text. [1][4][12]
- Do Claude marks identify the user or apply only inside the EU? refuted. Anthropic says the watermark and C2PA credential carry no user, organization, or chat identifiers. Marking is applied worldwide at launch because Anthropic says it does not yet have a durable way to scope it by region. [4][1][10]
What we found
- Anthropic’s Help Center policy is explicit about scope. New Claude models launched in the EU on or after 2 August 2026 support machine-readable marking at launch: generated text carries an embedded watermark, and generated files include C2PA Content Credentials where supported. Marks apply to output from supported models on Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and wherever Claude is offered, worldwide. Cloud-partner text from supported models on AWS, Google Cloud, and Microsoft Foundry is also watermarked. Older models sit in a legal transition period; Anthropic says it is adding marking to them and keeps a live model matrix on that page. [1]
- The change is dated, not merely a rewritten FAQ. The European Commission says Article 50 transparency obligations apply from 2 August 2026, lists Anthropic among Section 1 signatories, and reported about 190 organisations had signed by the end of July 2026. Anthropic’s news post is dated 14 August 2026 and was updated 1 September 2026 with detection-API information. Claude Platform release notes for 1 September 2026 state that text from Claude Fable 5.1 and Claude Mythos 5.1 carries Anthropic’s text watermark, with no request or response-handling changes required. The HTML model table on the Help Center did not convert into a complete per-model check grid, so this review does not claim a full current inventory of every older model. [2][3][4][5]
- The text mark is not a hidden Unicode character or a new API field. Anthropic describes a version of Google DeepMind’s SynthID-Text method, published in Nature in 2024: when the model has several equally good next-word choices, a keyed pattern in the sampling randomness leaves a statistical signature. Anthropic says nothing is added to the text, quality and price are unchanged, extra tokens are not required, and a reader cannot tell watermarked text from unmarked text. Independent coverage in The Verge and TechCrunch reported the same method and the EU-compliance motive. The Nature paper and Google’s SynthID docs add that detectability needs enough entropy and length, and that heavy rewrite or translation weakens the signal. [4][6][7][8][9]
- Files are marked differently, and not all files are marked. When Claude produces a supported image, video, or audio file in the code-execution sandbox, the Files API download carries a signed C2PA manifest. Anthropic’s docs say the credential identifies Anthropic as issuer, includes a timestamp, and records that Claude provided the file at a user’s request and may have created or modified it; it records nothing about the requester. Supported formats include PNG, JPEG, GIF, WebP, TIFF, HEIC, AVIF, SVG, MP4, MOV, MP3, WAV, FLAC, and M4A. Text files, PDFs, and office documents are not signed. C2PA is added where Claude’s file-generation path exists: Claude apps and Claude Platform (API), including Claude Platform on AWS and Claude in Microsoft Foundry. Re-encoding, conversion, screenshots, and metadata stripping remove it. [10][11][1][12]
- Detection is split. Anyone can use the in-browser checker at claude.com/check-content for C2PA on listed file types up to 100 MB; the file stays on device and the tool does not check text. Text watermark detection remains a private-preview API for regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations, EU civil-society groups, and enterprises that must verify watermarking for their own Act compliance. Anthropic says it plans to expand access and has not published Claude-specific false-positive or false-negative rates. [12][1][4]
- Anthropic’s own caveats matter as much as the headline. A detected mark means content may have been generated or processed by Claude, including proofreading, translation, or file conversion of human material. Absence of a mark does not mean the content is human: older unmarked models, heavy edit or paraphrase, very short passages, stripped file metadata, and unsupported features can all hide the signal. Code and tightly constrained factual text carry less watermark because there are fewer low-stakes word choices. Light editing may leave a mark; a full rewrite removes it. Anthropic says this does not change output ownership or user rights under its terms. [1][4]
What it means
Builders should treat supported Claude text as marked by default. There is no documented request parameter to turn the watermark off, marking requires no integration change, and Anthropic applies it at the model layer so a wrapper API still emits marked text. If you deploy Claude in a product, Anthropic tells you to assess Article 50 yourself: provider-side machine-readable marking is not a substitute for deployer labelling of deepfakes or AI-generated public-interest text. Do not promise customers that Claude output is unmarked, untraceable, or proven-human when a detector is silent.
Operationally, separate the two signals. Keep C2PA on generated media if your pipeline must retain provenance, and assume re-exports will drop it. Do not expect a public text detector, published error rates, or user-level attribution. For coding agents, the watermark is weaker in exact code than in comments and prose. For compliance, watch Anthropic’s live model matrix as pre-August 2026 models are backfilled, and do not treat a Help Center update as the start date—the legal clock started on 2 August 2026.
Still open
- Which older Claude models currently emit text watermarks on first-party versus each cloud partner, beyond the documented Fable 5.1 and Mythos 5.1 cases?
- When will the text detection API leave private preview, and what false-positive and false-negative rates will Anthropic publish?
- Does Google Cloud receive C2PA on generated files the same way Claude Platform on AWS and Microsoft Foundry do?
- Will Anthropic later add a regional off-switch, or keep global marking because it still cannot scope by geography?