Anthropic’s plan to watermark Claude’s writing has moved from announcement to product. Bruce Hopkins, an instructor who teaches for O’Reilly, wrote in O’Reilly Radar on 22 October that nine models already produce marked text: Mythos 5.1, Fable 5.1, Fable 5, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5, and Haiku 5.5. AI Insiders covered the August announcement and the criticism that followed. This piece adds how the finished mark works and what it means for people who publish Claude’s text.

Hopkins says the product surface makes no difference. Claude Desktop, Claude Code, and the API all emit the mark, so any code, email, blog post, newsletter, or white paper drafted with Claude carries it. He adds that anyone holding an AI watermark detector can tell the content was machine written.

The mark is built on SynthID-Text, which Google DeepMind released as open source. It hides in word choice. According to Hopkins, Anthropic says the system changes only the randomness that steers the model between near-equivalent words, so no hidden characters are inserted and no extra tokens are spent. That is why a paragraph pasted into Word, Slack, or a code editor keeps its signature.

The technique is called tournament sampling. For every word slot the model pulls several candidates and sends them through a knockout bracket, with a pseudorandom score deciding each round. That score is derived by mixing the last few tokens with a private key. Over hundreds of words, the text tilts toward high scorers in a fingerprint only the key holder can detect.

Hopkins is direct about privacy. Nothing about the user is encoded in the watermark, so a detector can say text is probably AI generated but cannot say whose Claude account produced it. For a freelancer or a company worried about exposure, that limits the damage: the mark flags the machine, not the person.

On quality, Hopkins points to Google’s test of roughly 20 million Gemini responses, where approval ratings moved by just 0.01 percent, a hair in favor of the watermarked model. That is a Gemini result, not a Claude one, and the August critics argued that thumbs-up rates cannot capture whether one synonym reads worse than another.

The law behind all this is the EU AI Act. Hopkins quotes Anthropic’s August statement that it and several other major providers are making the change to comply, and notes that since 2 August the EU has required AI providers selling into Europe to label what their systems generate. He cites the European Commission’s website for the claim that hundreds of companies have signed on, and adds that California’s SB-942 AI Transparency Act pushes the same direction.

Competitors are moving too. OpenAI says it is rolling out an invisible mark on qualifying ChatGPT and Codex text for European users, using a system it calls textGrain. By OpenAI’s own figures, its detector reaches 95 percent accuracy on passages of 400 tokens. Google has marked Gemini text with SynthID since 2024.

Detection is the open question. Hopkins describes public checkers for images, audio, and video, including Google’s SynthID.com and an OpenAI site, but the checker he names for text is simply “an AI watermark detector,” and he does not say who runs one for Claude. He also does not address how much editing erases the mark, beyond the fact that copying and pasting does not. Both gaps were open in August, and this piece does not close either.

Anyone publishing Claude text should assume it is marked and that someone with a detector can find out. The only published accuracy figure in Hopkins’s article, 95 percent at 400 tokens, belongs to OpenAI, so for Claude’s mark nobody outside Anthropic has shown how often a check would be right.

Reported by Bruce Hopkins for O’Reilly Radar on 22 October 2026.