Anthropic to launch watermark API letting third parties verify Claude-generated text

The technology is reportedly a variant of the SynthID Text method that Google DeepMind published in Nature in 2024. The underlying principle involves finely adjusting the randomness factor that comes into play as the model selects words one at a time, embedding a traceable pattern within the text. Anthropic explained that this process "does not affect the content, creativity level, or readability of text written by Claude." To a human reader, the sentences look no different from ordinary writing, but running them through the detection API can reveal the likelihood that Claude was involved.

Where it doesn't work

However, this method doesn't perform equally well across all types of writing. Short passages or sentences that simply list facts offer few alternative word choices, making it harder to embed a reliable pattern. The same applies to code. Text that has been purely proofread word-by-word by a human retains no watermark, since the word choices are the human's own. By contrast, Anthropic says translated text tends to retain the watermark well, since in translation, Claude ultimately chooses every word. Anthropic noted in its published FAQ that heavy rewriting of the original text can erase the watermark.

The watermark also only indicates the likelihood that Claude was involved in producing the text — it cannot distinguish whether Claude wrote the entire piece or only edited part of it. Nor can it determine whether a text was written by a human or by another company's AI.

A different approach from stylometric detection tools

Anthropic's approach is fundamentally different from existing AI-detection tools like Pangram. Services like Pangram don't have access to Anthropic's watermark key, so instead they statistically scan for AI-typical writing style or frequently used phrasing patterns. Watermark detection, on the other hand, confirms an actual trace left by Claude, which in principle can produce more reliable results.

EU regulation is driving global adoption

The EU AI Act is the backdrop for Anthropic's adoption of watermarking. In July, Anthropic signed the EU's Code of Practice on transparency for AI-generated content alongside roughly 190 other signatories. Since there's no practical technical way to restrict features by region, the watermarking applies simultaneously worldwide, not just in the EU. All Claude models released after August 2, 2025 support watermarking by default, and models released before that date will gain support gradually over the coming months. For non-text files, Anthropic uses the open standard C2PA, which attaches provenance metadata without altering the file itself.

Anthropic had previously announced on August 11 that it would make watermarking mandatory across Claude's outputs; this detection API represents the next step — opening up the ability to read those traces outside of Anthropic itself.

How it will work in practice

This API targets developers, not general users. A company that wants to add a feature to its service confirming whether text was generated by Claude would call Anthropic's detection API and receive a response indicating whether the watermark pattern is present. The exact launch date and application process haven't been disclosed in detail yet, but the likely use cases can be inferred. For example, schools or academic journals could use it as a first-pass filter to check for AI involvement in assignments or papers, and news or community platforms could use it to label AI-generated posts. Recruitment platforms could also use it as a reference indicator for whether a cover letter was AI-written.

Editor's take

Embedding a watermark and letting others verify it are two entirely different decisions. When Anthropic announced on August 11 that it would apply watermarking globally, that was a declaration of "we will leave a trace." This API is a declaration of "we will also hand others the key to check that trace." For Anthropic, this is a high-risk decision — it opens the door for outsiders to verify where and how its models were used, and if false positives or workaround cases pile up, it could actually erode trust rather than build it. Still, it's reasonable to assume that external pressure from the EU AI Act played a major role in this decision. Without regulation, there would have been little reason for Anthropic to voluntarily take on a transparency burden that competitors don't share.

OpenAI and Google have each experimented with their own methods of marking AI-generated text, but opening up detection authority itself to third parties has been rare. Google's SynthID has mostly been used within its own ecosystem, whereas Anthropic has taken that idea and opened it externally in the form of an API. Anyone who has worked with AI-detection tools in practice knows that stylometric detection tools have always struggled with false-positive rates. Watermark-based detection is inherently more accurate in principle, but the problem is that the exceptions — code, short text, and fully rewritten content — are exactly the cases most commonly encountered in real-world use. Ultimately, even once this API launches, a "not Claude" result should not be understood as definitive proof that a human wrote the text.

Domestic companies probably won't need to integrate this API right away, but content platforms and educational institutions should start considering adoption now. That said, it's safer to use watermark detection results as a supplementary indicator alongside existing stylometric analysis tools, rather than as standalone evidence. It's likely that within the coming weeks, OpenAI or Google will announce similar plans for externally accessible detection options. We've already seen this pattern play out multiple times — once regulation moves one player, the rest tend to follow.