AI
Aug 17, 2026
Anthropic is adding an invisible signal to Claude generated text that could make AI involvement easier to detect.
[For more news, click here]
For years, the debate around AI generated text has centered on a deceptively simple question. Can you tell when a machine wrote something? Anthropic is now trying to make that question easier to answer.
The company behind Claude has announced that new Claude models will generate text containing an invisible, machine detectable watermark. It will not appear as a visible label in the text. Instead, the watermark is built into the way Claude generates text, creating a pattern that can later be checked to see if Claude was involved. Anthropic is introducing the system as the EU's rules on AI generated content take effect, but the company plans to use it worldwide rather than only in Europe.
For people using Claude, the change may be difficult to notice. For the wider AI industry, however, it adds another approach to a problem that is becoming harder to ignore: how to establish where AI generated content came from.
How Claude’s Text Watermark Works
The watermark is created while Claude is generating a response rather than being added to a document afterward. When the model has several reasonable ways to continue a paragraph, the system can influence some of those choices according to a particular statistical pattern. Across a long piece of text, those choices can create a signal that can be detected with the appropriate key.
Anthropic's system uses an approach based on SynthID-Text, a watermarking technology developed by Google DeepMind. The company describes its watermark as unnoticeable to readers and says its testing found no meaningful effect on the quality or readability of generated text. The watermark is designed to remain with the text when it is copied and pasted elsewhere, and Anthropic states it may also remain detectable after some editing.
The most important distinction is between AI involvement and authorship. A successful detection does not prove that Claude wrote an entire document. Someone could write a report themselves and then use Claude to substantially edit it, for example, and a detection system could identify Claude's involvement without determining how much of the final work originated with the model.
The watermark also has technical limitations. Short passages provide less material making it difficult to identify the statistical pattern, while some tasks, including code generation, offer fewer opportunities for the system to influence word choices. For that reason, the technology is better understood as a provenance signal than as a definitive test for whether a piece of writing was produced by AI.
Why AI provenance is becoming more important
The watermark itself is only part of the process. There also needs to be a practical way to check for it. Anthropic is working on a detection API that will allow text to be tested for the watermark, although the company has not yet released all the details of how the service will work.
That could give organizations another way to monitor AI involvement in documents and other content as they establish policies around AI assisted work. Its usefulness, however, will depend on how reliably the mark can be detected across different types of writing and how well it remains detectable after editing. The system is also specific to Claude. Text produced by another AI model would not carry Claude's watermark, so organizations looking at AI generated material more broadly would need other ways to identify content produced by different systems.
Anthropic's announcement comes as governments and technology companies are looking for more reliable ways to identify AI generated material. The EU's AI Act is one of the main drivers, with rules requiring AI generated content to be detectable by machines from August 2, 2026. But the issue goes beyond regulation, generative AI is now producing large amounts of text, images, audio and video, while traditional ways of recognizing machine generated material are becoming less dependable.
Watermarking will not tell us whether the information is accurate or how much of the work came from a person. It can, however, indicate whether AI was likely involved. As businesses use AI more widely, that could make it easier to see where these systems were used and how they contributed to the final work.
Related Articles
How Anthropic and DXC Technology Are Moving Generative AI Into Mission Critical Enterprise Systems
Related Articles
