Anthropic extends hidden watermarking of Claude text to older models

Anthropic is extending its system of invisible watermarking for text generated by Claude. According to the Polish tech site Antyweb, such marking already works in the Opus 5.5 model, and on 3 November 2026 the rollout will cover the company's older models.

How the watermark works

Anthropic uses a variant of the SynthID-Text tool, developed by Google DeepMind. The model chooses words (tokens) in a way that creates a detectable statistical pattern across a longer passage. A reader will not notice it, but dedicated software can recognise it.

According to Antyweb, the mechanism does not add hidden characters or extra tokens and does not use up more of a user's limit. The company says it has seen no drop in quality. Some users on Reddit complain, however, that texts used to be better.

An estimate, not a verdict

The Antyweb author stresses that the detector only points to the possibility that AI was used. It is impossible to say with certainty that someone used artificial intelligence, or to establish how much of a text was written by a human and how much by a machine.

The method works better on longer texts, because there is more room for the signal. In the author's view, the only sure way to get rid of the watermark is to rewrite the text. Tools exist on GitHub, but their effectiveness is disputed. Translation may help, but it lowers quality. A single statement translated by AI may be enough for the whole text to be flagged. Conversely, a result of "100% human" does not rule out that the text was first generated and then rewritten.

Who can check

Anthropic is making its detector available for now in limited testing to media, researchers, universities and regulators. Access is due to widen over time. Ordinary users therefore cannot yet check texts on social media themselves.

Legal background

Antyweb links the change to the EU AI Act. Article 50 requires companies covered by it to mark AI-generated content in a machine-readable way. There is an exception for "standard editorial assistance" and actions that do not substantially alter the data. The rules have applied since 2 August 2026, and companies already on the market have a transition period until the end of the year.

Anthropic applies the watermark regardless of the user's location, meaning worldwide. According to Antyweb, OpenAI in ChatGPT complies only with EU requirements and does not use watermarking outside the EU.

What this means for you

The change affects Claude users worldwide, including in the UK, because the marking does not depend on country. People who use Claude for writing, for example letters, coursework, job applications or material for an employer, should know that longer texts may carry a hidden signal. For now, however, only selected institutions have access to the detector, and the result is an estimate, not proof.