Anthropic details Claude watermark system using Google SynthID tech

Anthropic has published technical details regarding its upcoming text watermarking system for the chatbot Claude, prompting widespread debate among subscribers and raising questions about how artificial intelligence-writing will be identified. According to Techcrunch reporting, the company issued a blog post on Friday seeking to answer some basic questions about how it will watermark the text generated by its chatbot Claude.

The move stems from an effort to comply with the European Union’s Artificial Intelligence Act’s Transparency Code, which requires AI companies to use systems that make it possible to identify AI-generated content. As outlined by Gizmodo, the announcement sparked an immediate backlash from a vocal segment of users on social media platforms who expressed frustration over privacy and professional use cases. Several high-profile accounts shared screenshots of cancelled subscriptions, citing the watermarking feature directly. Gizmodo noted that math and AI influencer John Ennis posted a screenshot of his cancellation, citing Anthropic's “ridiculous watermark idea” as a cause alongside other X users expressing anger.

Media additions

Despite reports of user revolt, Anthropic stated that in terms of the number of cancelations it’s seeing, there hasn’t been an increase since watermarking was announced. Some online complaints also referenced unrelated controversies, such as a recent Wall Street Journal article about CEO Dario Amodei’s wife’s past business practices as a cause for their cancelations.

The underlying technology relies on a version of the SynthID-Text approach that the Google DeepMind team outlined in 2024. Search Engine Journal notes that the implementation does not embed Unicode characters that are embedded into the text, nor does it rely on stylistic clichés like specific em dash frequencies or patterns that LLMs tend to use, like “It’s not this, it’s that” style of writing. Instead, the watermark operates by subtly altering the randomness used during token selection.

When Claude generates text, it frequently encounters low-stakes choices — like choosing between the words “overcast” and “grey” to describe the weather. Conversely, in situations with a single correct token, such as stating that Paris is the capital of France, the watermarking system is unlikely to kick in, because there’s only one “correct” token for that situation. Under the watermarking protocol, a secret key and a few words that come before to settle what word the model should pick guides these probabilistic choices. A human reader cannot distinguish the output from unwatermarked text, but a designated verification detector equipped with the proper decoding key can identify the statistical pattern. Anthropic also said that generating the watermark has a “negligible” impact on speed and token cost.

Technical specifications of the watermarking implementation include:

• Detection Method: Utilizes a version of the SynthID-Text approach that the Google DeepMind team outlined in 2024, verified via a watermark detection API.
• Impact on Output: Anthropic said that watermarking does not impact the quality of Claude’s output.
• Code Generation: Code, meanwhile, should have less of a watermark than other text, because the model will need to create working code and won’t have the freedom to choose between a variety of equally valid options, though arbitrary choices like comments within code can use the watermark.
• Editing Resistance: Light editing probably won’t remove the watermark completely, while a complete rewrite where every word is replaced will.

The system's behavior during human-AI collaboration remains a central point of discussion. Anthropic clarified that if a user writes a draft and has Claude proofread or edited by Claude, that will depend on the length of the text and how heavily Claude has edited it. If it’s only been lightly edited, nearly all the words will have been written by the human author and there’s very little (if anything) for the watermark to attach to.

Industry observers have noted that Anthropic will not stand alone in deploying these identification tools. Other major model developers have signed the same Code of Practice and will be implementing their own watermarks.