Anthropic's Claude Will Watermark Even Proofread Text Under EU AI Act
Anthropic will watermark supported Claude outputs worldwide under the EU AI Act, but broad labeling and easy removal could confuse content provenance.
Summary
On August 13, 2026, Anthropic said all new models worldwide will watermark supported output from launch under the European Union's AI Act, which covers AI generated or manipulated text, audio, images and video. It applies to models released after August 2 and gives earlier models until December 2026. Claude text, including Claude Code output, gets an invisible, machine-readable embedded mark; other files use signed C2PA provenance metadata where supported. Anthropic will mark any supported content Claude processes, including proofreading, translation, summarization and file conversion, though some platforms and features are incompatible.
This model-level system also tags grammar fixes and other standard editing that the law exempts when meaning is not substantially altered. Text marks bias word choices across a document and may survive copying or some editing, but another chatbot can erase them; screenshots, recordings and metadata tools can strip nontext provenance. Detection means only that Claude may have processed content, not generated it, while no mark does not rule out AI. Teachers and others could therefore mistake human work touched by Claude for wholly generated text.
Anthropic says marks do not change meaning, quality or readability and plans detection guidance and a text detection API, but provided no timetable or false positive and false negative tests; its statement to Ars Technica did not answer those questions or address exemption conflicts. AI Act section 50(4) generally requires no reader-facing label for AI written novels or marketing, and public-interest text escapes labeling if an identified, accountable editor reviews it. EU guidance frames transparency as protection against misinformation, manipulation, fraud, impersonation and consumer deception. Anthropic says other labs are taking similar steps; failures can draw fines up to 15 million euros or 3 percent of worldwide annual revenue.
Positives
- All new Anthropic models offered worldwide will watermark supported output from launch, extending compliance beyond the European Union.
- Claude Code output and other supported text will carry invisible, machine-readable marks that may persist through copying and some editing.
- C2PA metadata will provide digitally signed provenance for supported image, audio, video and other generated files.
- Anthropic plans a text detection API and technical guidance so users can check Claude marks themselves.
Risks & concerns
- Proofreading, translation and grammar corrections will receive the same Claude mark as wholly generated text despite the AI Act's standard-editing exemption.
- Another chatbot can erase text marks, while screenshots, recordings and metadata tools can remove nontext provenance.
- A detected mark only suggests Claude processed content, while an undetected mark does not exclude AI involvement.
- Text watermarking may substitute weaker word choices to preserve its statistical signal, although Anthropic denies effects on quality or readability.
- Anthropic provided no detection release date or false positive and false negative testing, and did not answer Ars Technica's exemption questions.
- Failed compliance could expose Anthropic to fines reaching 15 million euros or 3 percent of worldwide annual revenue.
