Starting this month, every piece of text generated or edited by Claude will carry an invisible AI watermark detectable with specialized tools. The change follows Anthropic’s decision to align with new European Union regulations requiring disclosure of AI-generated content, though the company is exceeding legal requirements.
The EU AI Act, which took effect earlier this month, mandates that providers implement measures to identify AI-produced material. Anthropic’s policy of watermarking all Claude output, including code, goes beyond those rules. Reactions have been divided. Some users have criticized the move as unnecessary, while others say it supports transparency as AI tools become more common in work and personal tasks.
The purpose of watermarking is to help users recognize AI-generated content. A job applicant using Claude to draft a cover letter might reconsider if employers can detect the mark. Similarly, employees receiving company memos could check whether the text was written by a human or an AI assistant.
However, the system has weaknesses. The EU AI Act includes exemptions that reduce its effectiveness. Computer code, for instance, is not required to carry watermarks. There is also an exception for “standard editing” and a vague clause allowing providers to skip watermarks if they are not “technically feasible.” That provision could lead to widespread opt-outs.
Related: Smart travel gadgets to pack for your next trip
Even when watermarks are applied, they can be removed. Claude’s system is built to withstand basic edits, but users can bypass it by processing the text through an open-weight model that does not add marks. Anthropic notes that the absence of a watermark does not confirm human authorship—it only means the content was not generated by a system that flags its output.
The EU AI Act and Anthropic’s policy show growing awareness of the need for AI transparency. Yet they also reveal the limits of technical solutions. Watermarks may not stop bad actors from producing low-quality, mass-generated content meant to manipulate or mislead. They could discourage casual misuse but are not a complete answer.
Other companies are taking different paths. Google and Meta have expressed support for the EU’s framework, while OpenAI calls its approach “layered,” focusing on content origins rather than universal watermarking. None of these methods solve the main problem: as long as AI tools remain widely accessible, some users will exploit them.
Watermarks encourage honesty. They will not stop AI-generated spam or misinformation. But they may make it harder to pass off machine-written text as human—at least until the next workaround appears.