Anthropic watermarks all Claude outputs globally from August 2026

Anthropic announced that it has signed the EU AI Act Code of Practice on transparency for AI-generated content and will begin watermarking all Claude outputs globally starting August 2, 2026. The watermarks will apply to all Claude products—the API, Claude, Claude Code, Claude Cowork, and Claude Tag—regardless of geographic location, even though the requirement stems from EU law. Anthropic is also working to retrofit existing models to support watermarking during a transition period the law provides.

The company will deploy two complementary marking methods. Text generated by Claude will carry an embedded, invisible watermark that preserves the text's meaning, quality, and readability and survives copying and pasting; the company states it "may persist through some editing." Because the watermark is applied at the model level, it will work across all Claude products. Files in supported formats—including .svg, .png, and .jpg images—will receive digitally signed provenance metadata based on the open C2PA standard, developed by the Coalition for Content Provenance and Authenticity. This signature proves Claude processed the file and can later reveal if the file has been tampered with. Text watermarks should also work through cloud partners such as AWS, Google Cloud, and Microsoft Foundry, though those platforms may not support the signed metadata layer. Anthropic plans to release verification tools to let users and third parties check the labels, though it has not specified a release date.

Anthropric is transparent about the system's limitations. A detected watermark does not prove Claude wrote the content; humans routinely use Claude for proofreading, translating, or summarizing existing material, so output may carry a watermark even though the underlying ideas came from a human. The absence of a watermark is equally inconclusive: the model may have shipped before watermarking launched, the text may have been heavily edited or translated after generation, the passage may be too short for reliable detection, or metadata may have been stripped through format conversion or screenshots. The real-world effectiveness of the watermarks will ultimately depend on how well they survive editing, reformatting, and translation—a test that has tripped up earlier attempts. Anthropic notes that third-party detectors could add support for its watermark, potentially giving them a more reliable signal than existing tools like Pangram, which use proprietary detection methods that do not reveal what triggered a result.

Anthropric is not alone in the watermarking race. Google DeepMind open-sourced its SynthID watermarking system and built it into Gemini models; SynthID tweaks token probability values during generation to create a watermark without degrading text quality and works across languages, though it struggles with edited text. OpenAI, by contrast, has possessed a text detector with 99.9 percent accuracy for roughly two years but has not released it, citing concerns that users can easily evade it through translation or rewriting, risks of stigmatizing certain groups, and likely worries that a public detector could damage OpenAI's own business. The issue is especially fraught in education, where unreliable detectors have already caused false cheating allegations, even as legitimate reasons exist to track AI use—studies show heavy reliance on AI can weaken critical thinking and writing skills, particularly among students who treat AI as a shortcut rather than a learning aid. Beyond academics, fraud has emerged: scammers are enrolling fake students at US colleges and using AI to complete coursework and collect financial aid. Anthropic's decision to watermark all outputs globally could also reshape its own business, since Claude is popular among students partly because older models produce natural prose; more reliable detection may make Claude less appealing to users seeking to bypass institutional oversight.