The new system is designed to identify Claude-generated text while raising questions about editing, detection, and AI-assisted work.
Anthropic has provided more details about its new watermarking system for Claude, explaining how the technology will identify text generated by its artificial intelligence models. The company says the watermark is designed to remain invisible to readers while allowing specialized detection systems to determine whether Claude was involved in producing the content.
Unlike a visible label or hidden characters inserted into a document, Claude’s watermark works through subtle statistical changes in the model’s choice of words. The system adjusts the probability of certain words during generation, creating a pattern that’s difficult for people to notice but can be detected using the appropriate cryptographic key. Anthropic said the approach is intended to preserve the normal quality and readability of Claude’s responses. The watermark is also applied at the model level, meaning it can follow content generated through different Claude products and services.
Minor edits may not remove the watermark.
One key question surrounding AI watermarking is whether users can easily remove the identifying signal by editing the text. Anthropic says the watermark is designed to withstand common changes such as copying, pasting and relatively minor alterations.
However, the company acknowledges that the system is not indestructible. A substantial rewrite can eventually weaken or eliminate the watermark, particularly when much of the original wording is replaced. This distinction matters because AI-generated material is often edited before publication, whether for accuracy, tone, grammar, or personalization.
Limitations with factual and technical content.
Anthropic also points to limitations when Claude is required to produce highly precise material. In areas such as factual answers and computer code, the model has less freedom to alter word choices without potentially affecting accuracy or functionality.
As a result, watermark detection may not always provide a definitive answer about whether Claude contributed to a particular piece of content. Instead, the system is intended to provide evidence of Claude’s involvement rather than prove that AI wrote every part of a document. Anthropic’s move comes as transparency requirements under the European Union’s AI Act take effect. The company has said its watermarking measures are intended to help meet regulatory expectations around identifying AI-generated material.
The broader development reflects a growing industry focus on AI content provenance. As generative AI becomes more common in workplaces, education, publishing and software development, companies and regulators are increasingly looking for ways to distinguish AI-assisted material from traditionally produced content.
