Anthropic has begun inserting an invisible watermark into outputs generated or edited by Claude, prompting pushback from some users.

TechCrunch reported on Aug. 12 local time that the move is intended to address transparency provisions in the European Union's AI law.

Under the new policy, Claude inserts an invisible code into sentences it writes or edits that can be identified by computer systems. It is not visible to the human eye, but is designed to allow machines to determine whether the output was generated by AI.

Some Reddit users objected that the move is excessive. One user said the watermark could become a tool that flags ordinary users more easily than skilled users who can bypass AI detection. The user said it could disadvantage students who used Claude to restructure paragraphs, office workers who summarised 200 pages of records, and users who asked for synonyms.

Another user called the watermark policy "unethical and disgusting" and argued that when users provide all instructions, context and the direction of revisions, Claude is only a tool. The user asked, "The prompt is entered by the user, so what exactly is Claude claiming credit for?"

Many users did not agree with that reaction. Some said the purpose of the watermark is not to claim authorship credit, but to put in place a minimal mechanism to identify AI-generated material.

One user said, "This is not about stealing credit," adding, "Since AI-generated material can cause various risks, there needs to be a way to identify it." Another user suggested there is little reason to take issue with a watermark unless the goal is to hide the fact that AI was used.

The controversy shows what conflicts a requirement to label AI-generated material can trigger in real-world use. Anthropic introduced a technical marking method to comply with regulations, but some users are raising concerns that traces could remain even when AI is used only as a supplementary tool in work or study.

Ultimately, the issue goes beyond the watermark itself to how far the boundaries of AI use should extend. Users are divided over whether using AI as a simple aid should be treated the same as submitting or posting AI-generated output largely as-is. The introduction of Claude's watermark is expected to keep debate going over how AI outputs should be labelled and where responsibility should lie.