Claude starts watermarking its text, and the part developers should watch is the detection API

Anthropic has started embedding an invisible watermark in the text its Claude models generate. The change is driven less by product strategy than by law: the EU AI Act’s transparency code took effect on August 2, and it requires providers to mark machine-generated content so other software can identify it. Anthropic is applying the watermark worldwide rather than only to European users, and on models launched from that date it is on by default with no opt-out. For developers who ship Claude-generated prose, the question is no longer whether the output can be identified as AI-written. It is who will be able to check, and when.

What Anthropic is actually doing

The method is an adaptation of SynthID-Text, a watermarking scheme first detailed by Google DeepMind in a 2024 Nature paper. When a model writes, it is constantly choosing among several next words that are all roughly equally good. SynthID-Text uses a secret key plus the words already written to steer which of those near-equivalent options the model picks, leaving a statistical pattern that a detector holding the key can recognize later. To a reader the result is meant to be indistinguishable from unwatermarked text.

Anthropic’s stated cost for this is close to zero. In its post, the company says watermarking “does not impact the quality of Claude’s output,” produces no extra tokens so “the model is the same price to serve and use,” and has a negligible effect on speed. Models launched on or after August 2 support watermarking at launch; for older models, Anthropic says it is still adding support during the law’s transition period. Anthropic frames the watermark as a property of what its models generate rather than a feature of one product surface, and its post does not carve out API traffic. The practical read for a team is that a content pipeline built on a current Claude model now emits marked text automatically, with no flag to turn it off.

Where the watermark holds, and where it doesn’t

The mark is not uniform across everything Claude writes. By Anthropic’s own account, code is barely watermarked because a model has little room to swap one valid token for another without changing behavior. Short passages detect poorly because there is not enough text to carry a reliable signal, and heavily factual writing is watermarked more sparsely for the same reason. Light editing leaves the watermark intact; only a full rewrite that replaces essentially every word removes it.

That distribution matters for how developers actually use Claude. The single biggest use case, generating and editing code, is effectively exempt, which is reassuring for anyone worried about marked source landing in a repository. The exposure sits with long-form prose: marketing copy, documentation, articles, support replies, anything a team drafts with Claude and lightly cleans up before publishing. The watermark is strongest exactly where AI authorship is most contested and weakest exactly where most developers spend their Claude budget.

The detection API is the part to track

The mark is inert without a reader. Anthropic says it “will soon be offering a watermark detection API,” with no launch date yet. That API is what would turn a private signal into an ecosystem tool: publishers, universities, plagiarism services, hiring pipelines, and content platforms could all query whether a given passage came from Claude. Because only Anthropic holds the key, it also becomes the single party that can answer that question, which concentrates the role of arbiter in one vendor. For a developer shipping customer-facing text, the concrete effect is that downstream systems may soon be able to flag Claude-written prose as machine-generated, and there is no supported way to opt that content out.

The pushback is about trust, not just mechanics

Not everyone accepts that the quality cost is zero. In a widely shared critique, John Gruber argued that steering word choice for any purpose other than the writer’s own meaning is a compromise by definition, and that a secret-key system “calls every single word choice into question.” Others flagged a subtler risk: a writer who uses Claude only to edit their own draft could later be flagged as having generated it. Anthropic’s position is that the sampling steer stays within equally good options and leaves output a reader cannot distinguish. Both claims can hold depending on how much the steer actually costs in practice, and independent testing has not yet measured that. The rollout beyond the EU is also a choice rather than a legal requirement outside Europe, which is part of why the reaction has been sharper than a compliance change usually draws.

What it means for developers

The working assumption should be that prose from a post-August-2 Claude model is detectable, that code largely is not, and that light editing will not strip the mark. If provenance matters to a project in either direction, wanting the detectability or wanting to avoid it, the levers are the model version and the detection API’s eventual terms, not an output setting. The wider frame is that the same EU deadline applies to every provider, so OpenAI and Google face the same clock. Anthropic is first to ship a detailed text watermark, which makes its approach the reference the others will be measured against.

What’s worth watching

The near-term signals are the detection API’s launch and terms, including who can query it and its false-positive rate, whether and when older Claude models gain watermarking, and any independent testing of whether the sampling steer measurably affects output quality. The broader signal is whether OpenAI and Google answer with comparable text watermarks under the same rule, or take a different route to the same compliance requirement. Stackmaven’s follow-up will revisit the detection API and any independent quality testing on or around the November mark.

Sources cited

1. Anthropic: How Claude's text watermarking works www.anthropic.com
2. TechCrunch: Anthropic shares more details about how Claude's new watermarks will work techcrunch.com
3. TechCrunch: Anthropic says it will watermark text generated by its AI models techcrunch.com
4. Euronews: EU compliance, delivered globally: Anthropic to watermark Claude's output worldwide www.euronews.com
5. Daring Fireball: Anthropic's watermark text adulteration in Claude daringfireball.net