Anthropic Explains How Claude’s New AI Watermarks Will Actually Work

SAN FRANCISCO — Anthropic has published a detailed explanation of how it plans to watermark text generated by its Claude AI, following an announcement earlier this week that triggered a wave of user debate — and reports of subscription cancellations — over what the feature means for people who use AI to help with work, school, and writing.

The watermarking system works by having Claude make subtle, invisible choices between equally valid words when generating text — for example, choosing “overcast” instead of “grey” to describe weather — in a pattern that is undetectable to readers but can be identified by anyone with access to a detection key. Anthropic confirmed the approach is based on the SynthID-Text system developed by Google DeepMind in 2024, and said it plans to release a watermark detection API that third parties can use to verify whether text was AI-generated.

The company was clear that the feature does not affect output quality. “To a reader, a watermarked response is indistinguishable from an unwatermarked one,” Anthropic said. The move comes as part of compliance with the EU AI Act’s Transparency Code, which requires AI companies to implement systems capable of identifying AI-generated content — and Anthropic noted Claude won’t be alone in this, as other major model developers have signed the same Code of Practice and will implement their own watermarking systems.

Several practical questions have emerged in online debate since the announcement. On whether someone could simply edit the text to remove the watermark, Anthropic said light editing probably won’t fully remove it, while a complete rewrite replacing every word will — “in the latter case, of course, it’s arguable whether the text can any longer be described as AI-generated,” the company said. On text that Claude only lightly proofread or edited rather than wrote from scratch, the watermark’s presence will depend on how heavily Claude contributed, since if nearly all words were written by the human author there is “very little for the watermark to attach to.”

Code generated by Claude will carry less of a watermark than prose, because functional code doesn’t offer the same freedom to choose between equally valid alternatives that written text does. Anthropic said watermarking could appear in code comments where arbitrary word choices are possible, but will have “a negligible effect on the actual code produced.”

The announcement has divided Claude’s user base sharply. Some users characterised the move as a conspiracy against users wanting privacy, while others argued the only reason to oppose AI watermarking is to deceive people about whether content is AI-generated. Business Insider reported dozens of users on X claiming to have cancelled their subscriptions in response, though the actual scale of cancellations is unclear.

Anthropic also drew a distinction between watermarking and the AI detection tools offered by companies like Pangram, which look for stylistic patterns and linguistic “tells” in writing to infer AI usage. Watermark detection is fundamentally different — it checks for a cryptographic pattern embedded in word choices, rather than trying to identify AI-like writing style. The two approaches can coexist and serve different purposes.