Anthropic says it will watermark text generated by its AI models
Anthropic has confirmed it will add watermarks to text generated by its Claude AI models, bringing it in line with new European Union transparency rules. The move follows the EU AI Act's Transparency Code, which came into force on 2 August 2026 and requires AI firms to label AI-generated or edited content so it can be identified by other systems, reflecting a broader push by regulators and platforms to make AI-generated material easier to detect.
All Anthropic models released after 2 August will automatically embed watermarking technology in both generated text and files, with files marked using the open C2PA standard; support for older models will follow. Anthropic says the text watermark is embedded at the model level, so it will appear regardless of which product it comes from—including the Claude API, Claude Code and Claude Cowork—and will persist even when text is copied, pasted or partially edited, though the company has not clarified exactly how much editing would remove it. The announcement follows similar steps by other firms, including Suno's watermarking of AI-generated music and Substack's partnership with Pangram to flag AI content, with Google, Meta, Microsoft, OpenAI and others also having committed to the EU code.
- Anthropic will watermark Claude-generated text and files
- Move complies with EU AI Act Transparency Code, effective 2 August 2026
- Watermark persists through copy-paste and some editing
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Supporters welcome the move as a responsible step towards transparency, arguing that as AI-generated text becomes harder to distinguish from human writing, watermarking helps readers, educators and platforms identify machine-generated content, curbs misuse such as plagiarism or disinformation, and demonstrates that AI developers can self-regulate rather than waiting for governments to impose rules. Extending the feature to older models is seen as a sign of good faith, showing that the commitment to accountability applies retroactively rather than only to new, more marketable releases.
The case against
Sceptics question the practical value of watermarking, noting that determined bad actors can often strip or evade such markers with simple edits, meaning the burden falls disproportionately on ordinary users while offering little real protection against misuse. Others worry about precedent: embedding traceable signals in text could enable unwanted surveillance of legitimate writers, chill experimentation with AI tools, or be seen as a symbolic gesture that reassures regulators and the public without solving the underlying problem of distinguishing human from machine authorship.