Anthropic published a blog post detailing how its new watermarking system for Claude-generated text actually works, aiming to answer questions that have circulated since the company first revealed the feature earlier this month. The rollout is tied to the company's obligations under the European Union's AI Act, whose transparency rules requiring AI providers to mark AI-generated content took effect August 2.
How the Watermark Actually Works
Large language models like Claude generate text one word at a time, weighing a list of potential candidates at each step and selecting the most sensible option based on the preceding text. Anthropic explained that its watermarking system takes advantage of moments where multiple word choices would work equally well — for instance, in a sentence like "The weather today was cold and…," several adjectives could follow with no meaningful difference in quality. In those cases, the system subtly biases which of the equally good options Claude selects, creating a statistical pattern that becomes detectable across enough text, without altering meaning, accuracy, or code functionality.
Anthropic said it is using a version of Google DeepMind's SynthID-Text approach, a technique published in 2024. Where an exact word or piece of code is factually required — meaning no alternative would work — the watermarking doesn't apply, since there's no equally valid choice to bias.
No Impact on Quality, Cost, or Speed
The company stressed that "watermarking does not impact the quality of Claude's output," stating that to a reader, a watermarked response is indistinguishable from an unwatermarked one — unlike traditional watermarks on banknotes or documents, which are visible to the eye. Anthropic said internal testing found no measurable effect on content, creativity, or readability, and cited Google DeepMind's own SynthID-Text research, which found no statistically significant difference in user ratings between watermarked and unwatermarked outputs when tested on a portion of Gemini traffic. The company also said the system won't increase the cost of running Claude and will have only a negligible effect on model speed.
What Triggers a Watermark — and What Doesn't
Because the watermark only applies to words Claude itself generates, the strength of the signal depends heavily on how much of a given text originated from the model. If a user submits their own writing to Claude purely for proofreading and Claude only makes light grammar or punctuation edits, there may be too few Claude-selected words for a detectable watermark to emerge. But if Claude is used for heavier rewriting, translation, or summarization, the resulting text can carry a detectable watermark even though the underlying ideas came from the user. Anthropic also noted that minor editing after the fact is unlikely to fully remove the watermark, while replacing the entire text would eliminate it. Very short passages are harder to assess reliably, since they contain fewer word choices for the system to draw a signal from.
Not a Standalone AI Detector
Anthropic was careful to frame the system as a provenance signal rather than a definitive detection tool. The company said the presence of a watermark "is not fully conclusive" that content was entirely AI-generated — only that it was processed by Claude at some point — while the absence of a detectable mark doesn't prove content wasn't AI-generated either, since older models, heavy editing, translation, or mixing with other text can all obscure the signal. Anthropic also distinguished its approach from third-party AI detectors like Pangram, which look for stylistic "tells" in writing (such as the construction "this isn't X, it's Y") rather than checking for an embedded statistical mark — the company said these are "fundamentally different" detection methods.
Files Get Metadata, Too
Beyond text watermarking, Anthropic said it is including signed provenance metadata, based on the C2PA standard, in other file types Claude generates or processes. A signed metadata label indicates a file was processed by Claude and can help detect whether the file was later tampered with. However, Anthropic acknowledged that images generated by Claude could still be screenshotted or otherwise stripped of this metadata, limiting its reliability in some contexts.
A Rollout Without an Opt-Out — and Some Pushback
The watermarking applies globally, without an opt-out, and Anthropic said it is working to extend the feature to older Claude models as well. The mandatory nature of the rollout has drawn criticism from some users, particularly lawyers, academics, and writers who use Claude to lightly edit or revise work that is substantially their own, and now face the prospect of that work carrying an AI marker they didn't choose. Critics have also raised concerns about the risk of misleading accusations — since even lightly-Claude-edited human writing can trigger a positive watermark detection — and about the company not yet publishing detailed accuracy thresholds or a dispute process for contested detections. Some users have reportedly cancelled Claude subscriptions in response, with debate playing out publicly on platforms like Reddit and X.
What's Coming Next
Anthropic said it plans to release a watermark detection API so users and third parties can check Claude's embedded watermarks and provenance metadata, though no firm timeline has been announced. The company noted that watermarking won't be unique to Claude — other major AI model developers have signed the same EU Code of Practice and are expected to roll out their own watermarking systems in response to the same regulatory requirement. For Anthropic's full technical explanation, see the official Anthropic blog post.
With regulatory deadlines already in effect and other major AI labs expected to follow suit, Claude's watermarking rollout looks likely to be an early test case for how the industry balances mandated AI transparency against user concerns over consent, accuracy, and control over their own writing.