Claude’s Invisible Watermark Marks Even Your Own Writing
Anthropic globally applies invisible watermarks to all Claude-generated and edited text to comply with the EU AI Act.
Stock photo for illustration only, not from the actual event
- Anthropic embeds invisible watermarks in all text generated and edited by Claude worldwide.
- The system operates at the token-selection level by biasing near-equivalent word choices.
- Impacts users who use Claude for proofreading, flagging human writing as AI-processed.
- This implementation aligns with Article 50 of the European Union AI Act.
Anthropic has begun embedding invisible watermarks into every piece of text produced by Claude, including text that users wrote themselves but asked Claude to proofread or polish. Announced in August 2026, this implementation is designed to comply with Article 50 of the EU AI Act, which requires providers of AI systems generating text to mark outputs so they can be detected as machine-made.
This technical measure is not restricted to the European Union; Anthropic applies the marking globally to output from supported models wherever Claude is offered. During generation, the model's token selection is subtly nudged to encode a hidden signal without altering the meaning, quality, or readability of Claude's response.
Stock photo for illustration only, not from the actual event
In an updated support article, Anthropic confirmed it has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. The framework details plans to embed watermarks in text and signed provenance metadata in generated files such as PDFs, images, and SVGs. The company has not publicly explained why it chose a global rollout over a regional restriction.
Anthropic's choice to enforce the watermark globally highlights a broader industry trend toward precautionary over-compliance with strict European regulations. By applying a universal standard, the company avoids complex regional carve-outs, though it inadvertently shifts regulatory burdens onto everyday users outside the EU who utilize AI strictly as an editing assistant.
The underlying mechanism relies on statistical text watermarking. Language models select tokens from a probability distribution at each step. Rather than normal randomness, the watermarking technique uses deterministic, key-dependent selection among near-equally probable tokens to encode bits of the watermark signal, which remains detectable by anyone holding the decryption key.
"the very fact that there is generally no ‘best next token’ with 100% certainty is precisely why the trick works."
Hacker News commenter
The mathematical reality diverges from the user experience when Claude is used for light editing. The model regenerates the entire response token by token rather than applying isolated corrections. Even when the chosen word matches what the user originally wrote, the model selects it through the biased probability distribution, imprinting a statistical signature.
Consequently, downstream systems like university plagiarism checkers, publisher filters, or recruitment screening tools will flag watermarked text as AI-generated, regardless of whether a human authored 90% of the words. A watermark that cannot distinguish between AI authorship and human writing polished by AI labels contact rather than true authorship.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment