Anthropic is putting an invisible watermark into text generated by Claude. You can’t see it. You can’t point to a character or a hidden tag. And right now, almost no one outside Anthropic can reliably check for it.
That has left a lot of people confused, irritated, or quietly wondering if the model is now quietly changing the words it chooses. Here’s the plain-English version of what’s going on.
What the watermark actually is
Large language models don’t write sentences the way humans do. At every step they look at a list of possible next words and pick one. When several options are roughly equally good—“overcast” versus “grey,” “stream” versus “brook”—the choice is normally settled by randomness.
The watermark changes the source of that randomness. Instead of pure chance, Claude uses a secret cryptographic key combined with the words that came before. Over a long enough stretch of text those small, low-stakes choices create a statistical pattern. Anyone who has the matching key can test the sequence of words and estimate the probability that Claude was involved.
Nothing is added to the text. There are no hidden characters, no special spacing, no metadata you can strip in a text editor. The signal lives in which words were selected when the model had a real choice.
Anthropic says this is a version of Google DeepMind’s SynthID-Text technique. They also say internal testing and the original research found no measurable drop in quality, creativity, or readability. The words still have to be ones the model already considered reasonable. It will not turn “aircraft carrier” into “dinghy.”
Why you can’t just “look at” it
Because the pattern is statistical and keyed, you need the private key (or a detection tool that has access to it) to see the signal. Anthropic has said it will release a detection API. As of mid-August 2026 that API is still “coming soon.” When it arrives, it will report a probability or confidence score that Claude was involved, not a simple yes/no.
Light editing or pure proofreading of human-written text usually leaves little signal, because most of the words remain the original author’s. A thorough rewrite or a full paraphrase through another model can erase it. At that point, of course, the text is no longer the original Claude output.
The sycophancy concern
Some people (myself included) have asked whether this is just another subtle pressure on token selection—similar to the way sycophancy already nudges models toward agreeable but lower-value phrasing.
The technical distinction is real. Sycophancy changes the model’s preferences: it shifts the ranking of what counts as a good answer. The watermark does not. It only re-seeds the random draw among options the model already ranks as roughly equal. The underlying ranking of accuracy, relevance, and quality is supposed to stay the same.
Whether that distinction holds up under real high-stakes work is a fair open question. The published tests are mostly aggregate quality ratings, not deep stress-testing on precise technical or scientific writing. Any systematic interference with token choice is worth watching. But the design intent is to keep the interference small and confined to near-ties.
Why this exists at all
Blame the European Union.
Article 50 of the EU AI Act requires providers of generative AI systems to mark synthetic text, images, audio, and video in a machine-readable way so it can be detected as artificially generated. The obligation took effect in August 2026 for new models; older models have a transition window until early December. Anthropic, along with Google, OpenAI, Meta, Microsoft, and roughly 190 other organisations, signed the related Code of Practice. Because Anthropic says it does not yet have a reliable way to apply the mark only inside the EU, it is rolling the watermark out globally.
The stated policy goal is transparency: reduce the risk of large-scale deception, misinformation, and loss of trust in the information environment. Whether a company-specific statistical signal that can be washed out by paraphrasing actually delivers that goal is debatable. The effort is real. The practical payoff, at least so far, looks limited.
Should you stop using Claude?
Not because of the watermark itself.
The watermark does not identify you, your organisation, or your specific chat. It only answers (with a probability) whether Claude was likely involved in producing or heavily editing the text. It does not transfer ownership or create new legal liability. Quality is claimed to be unchanged.
What it does introduce is a new, imperfect provenance signal that platforms, employers, or regulators may eventually start checking. If you need text that carries no machine-detectable Claude fingerprint, you will need to rewrite it. That has always been true of any model once third-party detectors improve; this simply makes the signal more systematic for one provider.
Other major labs are moving in the same direction under the same regulation. Google is using SynthID in Gemini. xAI has not signed the same Code of Practice and has indicated it is not embedding a statistical text watermark of this type. The broader trend is clear: machine-readable marking of AI output is becoming the new baseline in regulated markets.
The honest summary
Claude’s watermark is a regulatory compliance measure, not a quality feature and not a user-tracking system. It lives in the random choices among good words. You cannot see it with your eyes. Detection tools are still rolling out and will give a probability score, not a binary answer. Heavy rewriting removes it. Light proofreading of human text usually leaves little signal.
It is not a reason to abandon Claude if the model otherwise works for you. It is a reason to understand exactly what the mark can and cannot prove, and to treat any claim of “definitive AI detection” with the scepticism it still deserves.
The EU required the industry to do something. This is the something they chose. Whether it was worth the complexity remains an open question.