Anthropic Adds Invisible Watermarks to Claude-Generated Text

Anthropic has started watermarking editorial text generated by Claude, inserting invisible code intended to identify the material as AI-generated. The policy is linked to the European Union AI Act’s Transparency Code, which requires technology companies to label AI-generated or AI-edited content in a way computer systems can recognise.

The change has prompted a visible backlash among some Claude users on Reddit. Critics argued that watermarks could expose people who use the chatbot for school or workplace tasks, including rewriting, summarising and drafting. Other participants in the discussion supported the approach, saying detection is relevant where AI-generated output could be presented to others without disclosure.

Machine-readable identification

Anthropic’s stated direction is to mark Claude’s editorial output through invisible code rather than a visible label in the text. That design targets automated identification of generated material, matching the Transparency Code requirement that AI-created or AI-modified content be identifiable to computer systems.

The dispute highlights the difference between using a tool during a workflow and submitting its output unchanged. A summary of a long transcript can be useful research support, but copying it verbatim into published work would raise separate editorial and ethical concerns. Similarly, a student or employee who pastes generated prose directly into a submission may face rules that differ from those governing brainstorming or revision assistance.

Debate over disclosure and responsibility

Some Reddit posters framed the watermark as an unfair restriction on users who supply prompts, context and refinements. Others replied that the marker is not a claim of credit by Claude, but a way to detect AI-generated output in situations where undisclosed automation may matter.

Anthropic’s move follows Anthropic’s Claude text watermarking initiative as the company develops machine-readable labelling for Claude text under the EU transparency framework. The continuing debate also includes concerns about the training data used by frontier models, although those concerns are distinct from the question of whether generated output should be detectable.

For organisations, the immediate implication is operational rather than rhetorical: teams using Claude should define when AI assistance must be disclosed, require human verification of factual work, and avoid treating generated text as untraceable final copy.