Anthropic will begin applying an invisible text watermark to output from four more of its Claude models on Tuesday, November 3, extending a marking system its newer models already carry. The change covers Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6, and it will roll out worldwide rather than in Europe alone. Users will see nothing different on screen, cannot opt out, and need not change any settings, according to the customer email describing the rollout.
What changes on November 3
Until now, the watermark has covered Anthropic’s newer models, including Claude Opus 5.5, Opus 5, Sonnet 5.5, and several others launched after August 2. The four older models joining on November 3 are among the company’s most widely used earlier models, including through the Claude API, Claude Code, and cloud partners such as Amazon Bedrock.
The November 3 date comes from Anthropic’s customer email, sent on Friday, October 30. Anthropic has not yet published that date on its public help page, whose table still leaves the text-watermark columns blank for the older models. The rollout on Amazon Bedrock may take a few additional days.
Three older models are not in the November 3 list: Claude Opus 4.5, Sonnet 4.5, and Haiku 4.5. The help page still says the company is working to add marking support for earlier models, suggesting further dates will follow. Sonnet 4.5 leaves the Claude API on November 30 anyway, so its absence may simply reflect that retirement.
How an invisible mark works
When Claude writes, it often faces choices between words that fit equally well, and normally a random number settles the choice. With watermarking, a secret key combined with the preceding words settles it instead, leaving a statistical pattern across a long passage that only a detector holding the key can test for.
The company says nothing is added to the text, there are no hidden characters, no extra tokens are consumed, and quality and readability are unchanged. Anthropic’s approach is based on SynthID-Text, the method Google DeepMind described in a 2024 Nature paper. It says the mark cannot be traced to a specific person, organisation, or chat.
Because the mark lives inside the text itself, it can travel with copied and pasted passages and may survive some editing. Code generally carries less of the mark than ordinary prose, because code offers fewer free word choices for the pattern to hide in. Short passages give too little signal for a reliable test, for the same reason.
What the mark cannot prove
It cannot distinguish between text Claude wrote outright and text Claude heavily edited, translated, or summarised for a person. A detected watermark can only show that Claude was likely involved with the content at some point, the company says. Anthropic is unusually frank about the limits of its own system, and those limits matter more than the rollout date.
A missing mark is therefore not proof that a person wrote something, and a present mark is not proof of authorship. A passage a person wrote and merely had proofread by Claude could carry the mark, while a passage Claude wrote and a person then rewrote completely could carry none. Light editing probably will not remove it, Anthropic says.
Who can actually check is another limit. Anthropic’s detection API is in private preview for groups it is required to serve under European law, including regulators, law enforcement, media, fact-checkers, researchers, and educational organisations. The general public cannot paste text into a Claude checker today, unlike Google, which opened its SynthID detector to everyone in October.
Our read: compliance is driving a quiet default
Models launched before August 2 were given a transition period, which is why the older models are being added later than the newer ones. Anthropic says it is adding watermarking to comply with the European Union’s AI Act, having signed the EU Code of Practice on Transparency of AI-Generated Content in July.
The global rollout is the telling detail. Anthropic says it cannot yet restrict the feature to specific regions in a durable way, so a rule written for Europe now applies to writing produced everywhere, including the United States. OpenAI is taking a similar path in Europe, so this is becoming an industry default, not a single company’s experiment.
xAI is taking a comparable step with a hidden watermark for Grok API text on November 7, which suggests that invisible marking is becoming the standard answer to provenance, whether or not readers asked for it. That standard still leaves the hardest question open: who gets to check, and what a positive result is allowed to mean.
There is no way for a writer to prove a negative under this system. A mark that only regulators and researchers can test, and that cannot separate writing from editing, gives institutions a tool while leaving ordinary readers exactly where they were. The companies building these systems have answered compliance first, and public trust second, if at all.
Why It Matters
Plenty of apps, businesses, and writers still run on Opus 4.7, Opus 4.6, and Sonnet 4.6, and from Tuesday what those models write becomes checkable by the organisations Anthropic lets in. Publishers, schools, and employers that rely on Claude output should expect a detectable trace, even when a person did most of the work and Claude only polished it.
The mark will shape how people write with Claude long before it settles who actually wrote what. The strategic question is harder, because a watermark that survives copying but not rewriting rewards heavy editing over honest disclosure. Enterprise watermarking tools such as WRITER’s offer companies their own configurable marks, but Anthropic’s version offers no such choice to the writer.
The practical advice in Anthropic’s email is simple: customers need not change anything unless they rely on the watermark for their own compliance and use non-default sampling settings. For everyone else, the change is invisible by design, which is precisely why its limits deserve to be stated plainly before anyone treats a detection result as a verdict.
Background
Anthropic explained the watermark method in a technical post in August, and its help page on how Claude marks AI-generated content was updated on October 28. For images and other supported files, Claude attaches content credentials through C2PA metadata instead of a text watermark, a separate system with its own public file checker.
Source: NDTV Profit report and Madrobot analysis
0 Comments