In August 2026 Anthropic announced that future text outputs from Claude models will include a watermark. According to the company, the watermark provides a way to estimate the likelihood that Claude contributed to producing a given text. Anthropic and several other major AI providers are making this change to comply with the EU AI Act; the EU requirements for providers serving its market came into effect on August 2, 2026.
Which Claude models are affected
At the time of Anthropic’s announcement, the following Claude models were listed as having AI text watermarks on all generated outputs:
- Claude Mythos 5.1
- Claude Fable 5.1
- Claude Fable 5
- Claude Opus 5.5
- Claude Opus 5
- Claude Opus 4.8
- Claude Sonnet 5.5
- Claude Sonnet 5
- Claude Haiku 5.5
The watermark appears across interfaces — Claude Desktop, Claude Code and the API — so code, emails, blogs, articles, newsletters or white papers produced with Claude will carry the watermark and can be identified by anyone with a watermark detector.
Cryptography vs. steganography
It’s useful to distinguish cryptography from steganography in this context. Cryptography hides content by making it unreadable without a decryption key (as with HTTPS). Steganography, by contrast, leaves the text in plain sight but embeds a hidden signal that is only visible to those who know how to look — similar to a security thread or watermark on a banknote.
Anthropic’s watermark uses a steganographic approach: the signal is embedded in word choices inside otherwise ordinary-looking text, so the watermark persists when the text is copied or moved.
How the watermark works technically: SynthID-Text and tournament sampling
Anthropic’s watermarking builds on SynthID-Text, an open-source project from Google DeepMind. The method does not add hidden characters or extra tokens; instead it changes the source of randomness the model uses when choosing among near-equivalent tokens.
SynthID-Text uses a technique called tournament sampling. At each token position the model draws several candidate tokens and runs a knockout-style tournament. Winners are chosen based on pseudorandom g-values derived from hashing the previous tokens together with a secret key. Over the span of hundreds of words, token choices show a slight bias in a pattern visible only to someone with the key; this pattern is not apparent to a human reader. Anthropic describes the effect as a “slight favoring” of certain tokens, which typically does not alter sentence meaning.
What the watermark does not reveal
Anthropic emphasizes that the watermark does not insert hidden characters, does not consume extra tokens, and does not carry information about the user. A watermark detector can establish that a text was AI-generated, but it cannot identify which Claude account created it.
Other providers and the regulatory landscape
Because the EU mandate affects multiple providers, other major companies have adopted watermarking as well. OpenAI announced support for watermarking eligible ChatGPT and Codex outputs in the European Union and calls its approach “textGrain.” OpenAI published test statistics indicating their watermark detector reaches 95% accuracy for passages of about 400 tokens, suggesting only a few paragraphs can be sufficient for detection.
The requirement for watermarking is not limited to the EU: for example, California’s SB-942, the California AI Transparency Act, also requires watermarks on AI-generated content. These regulatory regimes cover not only text but also images, audio and video.
Google implemented SynthID Text watermarking in its model outputs as early as 2024. Google and other companies — including OpenAI, Nvidia and Kakao — have published public tools to check non-text media for watermarks; SynthID.com is an example that can analyze images, audio and video.
Practical benefits of watermarking
Watermarks can be a practical tool for assessing authenticity and mitigating malicious uses of AI. The article cites a real-world example: after a startup announced venture-capital funding, scammers impersonated the CEO to request bank transfers. In that case, suspicious email domains gave away the fraud. But with advanced tools such as voice cloning, attackers could use AI to impersonate executives by phone; reliable watermarking on audio or other media could help detect such attacks.
Conclusion and practical advice
Major LLM providers are incorporating watermarking into AI-generated content and, given regulatory pressure, this practice is likely to remain. Watermarks are a useful complement for determining whether a piece of content — text, image, audio or video — was produced by AI, especially when combined with detectors and public checking tools.
For those who want to see live demonstrations of how AI watermarking is applied and detected, the author points to the O’Reilly course Understanding AI Text Watermarking, which includes live examples.
