Anthropic announced that Claude's output will carry an imperceptible text watermark and signed C2PA provenance metadata for files, worldwide, to comply with the EU AI Act's Article 50 transparency rules. Testing on August 13, 2026 against Claude Opus 5 and Claude Fable 5, on both the desktop app and the API, found no C2PA provenance metadata on any generated image files, confirming the file-marking mechanism has not yet been activated. Anthropic's documentation states only models launched on or after August 2, 2026 support marking at launch, and no new models have shipped since then; older models are pending a rollout with no timeline. The piece also debunks four circulating myths: that the watermark is strippable hidden Unicode characters (it's actually a statistical pattern woven into word choices, matching Google DeepMind's SynthID-Text approach), that anyone can currently detect Claude text (no detection tool exists yet, even for Anthropic's own customers), that the watermark identifies the specific user or account (it's a model-level signal, not a per-customer marker), and that editing either always or never removes it (text watermarks degrade with heavy rewriting, while C2PA file metadata is stripped by re-rendering or screenshotting but otherwise persists).
Table of contents
IntroductionWhat we could actually verifyWhat Anthropic actually committed toMyth 1: it’s hidden characters you can strip outMyth 2: anyone can now detect Claude textMyth 3: it identifies youMyth 4: editing removes it, or nothing doesWhat this means for your businessConclusionQuestions this post answers
Does Claude currently add C2PA provenance metadata to the images it generates?
No, as of August 13, 2026, images generated with Claude Opus 5 and Claude Fable 5 through both the desktop app and the API carried no C2PA manifest at all, checked with the c2patool command-line utility which returned 'No claim found' every time. Anthropic's own documentation says only models launched on or after August 2, 2026 support machine-readable marking at launch, and no such models had launched yet, while marking for existing models is still in progress with no announced timeline. daily.dev surfaces this kind of hands-on verification for teams tracking AI vendor compliance claims before they rely on them.
Can Claude's text watermark be removed by stripping hidden Unicode characters?
No, because Anthropic's text watermark is not made of hidden characters at all. It works as a statistical pattern embedded in the model's word choices, similar to the approach in Google DeepMind's SynthID-Text, so removing invisible characters does nothing since there is nothing separate to strip; the watermark is the text itself, only detectable in aggregate by a verifier. Developers evaluating AI content provenance tools follow deep dives like this one on daily.dev.
Can anyone detect whether text was generated by Claude and watermarked?
No, no third party, including Anthropic's own customers, can currently detect Claude's text watermark because Anthropic has not released a detection tool. Anthropic states it is working to enable users and third parties to detect embedded watermarks, with technical documentation still forthcoming. Existing public AI detectors only guess based on surface writing features and do not read the actual watermark signal. Teams weighing AI disclosure risk keep up with watermark and detection developments through daily.dev.