Claude Now Marks Its Own Writing — Here's What That Means

On August 14, 2026, the company behind the AI assistant Claude explained how it will start marking the text that Claude produces, so that people can later tell it was written by AI. The method is based on a technique called SynthID-Text, which was originally created by Google DeepMind in 2024.

The reason for this is a new law. The EU AI Act went into effect on August 2, 2026, and it requires AI companies to make it possible to identify content generated by AI. Claude models released on or after that date must carry the mark. Other major AI companies have agreed to the same rules and are expected to add their own versions.

The way it works is surprisingly simple in concept. When Claude writes something, it constantly chooses between words that mean roughly the same thing — like picking "overcast" or "grey." The watermark quietly nudges those choices in a particular pattern, like a faint watermark hidden inside a piece of paper. You cannot see it by reading the text. But if you have the right digital key, you can scan the text and confirm that Claude wrote it. Anthropic says the watermark does not affect the quality of the writing.

Anthropic plans to release a tool — called a detection API — that will let other companies check whether a piece of text came from Claude. This is different from existing AI-detection tools, like those made by a company called Pangram, which try to guess whether something was AI-written by looking for patterns in the style of the writing. Checking for a watermark is a direct test, not a guess.

How long the watermark lasts depends on what happens to the text afterward. Small edits will probably not erase it. But if someone rewrites every single word, it will be gone. If Claude only fixes a few words in something a person already wrote, there is not much for the watermark to attach to. Computer code is different: because code has to actually work, there is less room for the model to make arbitrary choices, so code will carry less of a watermark than regular writing. Anthropic says the effect on the code itself is negligible.

The broader context is a tension between tracing where content came from and people's desire for privacy. Business Insider reported that "dozens" of Claude users on X said they canceled their subscriptions over the watermark. Some users worry that watermarks could reveal that they used AI in situations where they would prefer to keep that private.

There is a gap worth noting between what the law assumes and how the technology actually works. The EU AI Act assumes that AI-generated content can be reliably told apart from human writing. But SynthID-Text is a probability-based system, not a simple yes-or-no stamp. It works well on long, unedited text and gets weaker on short passages, heavily edited text, and code. The detection tool will give a yes-or-no answer, but how sure it is underneath that answer will depend on how long the text is and how much it has been changed. Anyone relying on this for compliance should understand that the watermark fades gradually, not all at once.

Anthropic's recent product timeline gives context for the timing. The company released Claude Sonnet 5 on June 30, 2026 and Claude Opus 5 on July 24, 2026. Both came out after the EU AI Act's August 2 deadline, so both must carry the watermark. Anthropic also announced on August 4, 2026 that Mariano-Florentino (Tino) Cuéllar will join as Chief Global Affairs Officer, a hire that fits with the company's growing involvement in regulatory discussions in the EU and beyond.

The SynthID-Text technique is not new — Google DeepMind published it in October 2024. But Anthropic adopting it is the first time a major AI company has used a watermarking standard developed by a competitor. If other companies that signed the EU AI Act's rules adopt compatible approaches, the detection tool could become a shared way to verify where AI text came from, across different AI providers. That is possible, though Anthropic has not said whether its watermark or detection tool will work with other companies' systems.

For anyone using Claude through its programming interface, the key facts are these: text generated by Claude will carry a hidden signal that survives moderate editing but not a complete rewrite. Code will carry very little signal. A detection tool is coming that will let others check for the watermark automatically. How useful the watermark is in practice will depend on how much of your text is long and unedited versus short and heavily changed.