Anthropic released a blog post on Friday explaining how its new watermarking system for Claude works. The company said it wanted to answer common questions after announcing the change earlier in the week.
The watermark is designed to help identify when text was generated by Claude. It works by taking advantage of moments when the model has more than one equally good word choice.
For example, if Claude is writing a sentence about the weather, it might choose between the words "overcast" and "grey." Both work equally well. Anthropic says the watermark uses a hidden key to guide which of these low-stakes choices Claude makes, creating a pattern that can later be detected.
How the Watermark Affects Output
Anthropic says the watermark has no effect on the quality of Claude's answers. A reader cannot tell the difference between watermarked and unwatermarked text.
The company pointed to internal testing and a study from Google DeepMind, whose SynthID-Text method Anthropic is using. That research found no meaningful difference in user ratings between watermarked and unwatermarked responses.
Anthropic also said the process adds no extra tokens. This means watermarking will not slow down Claude or make it more expensive to use.
The watermark does not add any hidden characters or extra information to the text. It also does not contain any details that could identify a specific user, organization, or chat.
Why Anthropic Is Making This Change
The update is tied to the European Union's AI Act. As of August 2, companies serving the EU market are required to mark AI-generated content.
Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. Other major AI companies signed the same agreement and are expected to release their own watermarking systems.
Anthropic said it applied the watermark globally at launch. The company explained it does not yet have a reliable way to limit the watermark to only the EU market.
The watermark works less effectively on short pieces of text. Anthropic said longer passages give more chances for the pattern to appear, making detection more reliable.
Factual writing and code contain fewer watermark opportunities. This is because there is often only one correct answer, leaving no room for a low-stakes word choice.
Anthropic said comments within code can still carry a watermark, since word choice there is more flexible. The actual code itself is barely affected.
The company also addressed whether editing could remove the watermark. Anthropic said light editing likely will not erase it, but a full rewrite of every word probably would.
Anthropic noted this raises the question of whether a fully rewritten text can still be called AI-generated at all.
Reaction to the announcement has been mixed. Some users argued the watermark helps with transparency, while others online called it invasive. Business Insider previously reported that some users said they had cancelled their Claude subscriptions in response.
Anthropic said it plans to release a watermark detection API. Details on how that tool will work are still being finalized.
The company also confirmed it is working to add watermarking to older Claude models. This rollout is expected to happen gradually over the coming months.
