Claude's newest models now hide a detectable watermark in the text they write

What Anthropic announced

As oreilly.com reported, Anthropic announced in August 2026 that its Claude models will produce text containing a watermark. The purpose is to let someone estimate how likely it is that Claude helped write a given passage. Anthropic says the change is part of complying with the EU AI Act, and that other major AI providers are making similar moves.

The article notes that, as of August 2, the EU expects AI providers serving its market to mark AI-generated content. Other big developers have signed the same Code of Practice, so each will use its own watermark.

The affected list is long. It covers models across the Opus, Sonnet and Haiku families, plus Mythos 5.1 and the Fable 5 and 5.1 models. It does not matter how you reach Claude: the desktop app, Claude Code and the API all produce marked text.

Hidden in plain sight, not locked away

The author separates two ideas that people often mix up. Cryptography scrambles a message so only someone with the key can read it; it is what protects your card number on an HTTPS website. Steganography leaves the message fully readable but tucks a second signal inside it, for people who know where to look.

The paper-money comparison fits: hold a large bill up to the light and a watermark and security thread appear. Claude's watermark is the steganography kind. Nobody is blocked from reading your text. It simply has an extra layer that a detector can check.

How the word-choice trick works

Claude's approach is based on SynthID-Text, an open source project from Google DeepMind. When a model writes, it often faces several words that mean nearly the same thing. The watermark quietly steers those picks in a pattern that is invisible when you read but measurable by software holding the secret key.

The technique is called tournament sampling. At each step the model pulls a few candidate words (technically, tokens, which are small chunks of text) and puts them through a small elimination contest. Scores for the contest come from the few preceding words combined with the secret key. Across hundreds of words, the text leans a little toward high-scoring choices, and that lean is the fingerprint.

According to Anthropic, only the source of randomness changes. No hidden characters are inserted, no extra tokens are used, and the mark holds no information about you. A detector can say the text looks AI-generated, but not whose account made it. Because the signal lives in the words themselves, it stays with the text when you move it into a document, email or chat message.

The author adds that the nudge is small enough not to harm quality. In a Google test on about 20 million Gemini responses, user approval barely moved, with the watermarked version very slightly ahead.

Everyone is doing it, and why

Anthropic is not alone. The article says hundreds of companies have agreed to follow the EU's approach, and that OpenAI has announced an invisible watermark for eligible ChatGPT and Codex text in the European Union, called textGrain. OpenAI reports its detector is 95% accurate on passages of 400 tokens, which suggests a few paragraphs are enough to detect.

Google has marked its text with SynthID since 2024. It also runs a public website, SynthID.com, where anyone can check images, audio and video for watermarks from several providers, including OpenAI, Nvidia, Kakao and Google. OpenAI has launched a similar site for non-text media. California's SB-942 AI Transparency Act adds a further requirement, covering text, images, audio and video.

The author argues this is useful rather than sinister. He describes a scam attempt at a startup where fraudsters posed as the CEO to request bank transfers. That one was easy to spot, but cloned voices could make a similar scam far more convincing, and reliable marks on AI media would help people judge what is real.

What to do about it

The practical lesson is that AI-written text should be treated as potentially identifiable. That is not a reason to avoid Claude. It is a reason to be open about using it, especially where rules apply, such as schools, job applications and published work.

• Teams: agree on when you tell customers AI helped draft something.
• Students and job seekers: check the stated rules, and rewrite in your own voice.
• Creators: edit heavily and label AI help honestly.

One caution: the source describes the detection claims mostly from company statements and the author's own testing of a media site, so treat accuracy figures as vendor-reported.