Anthropic adds invisible watermark to Claude AI texts to comply with EU AI Act

Anthropic introduces invisible watermark for Claude

The US company Anthropic has announced that texts from its chatbot Claude will carry an invisible, machine-readable watermark to comply with EU requirements for AI-generated content. The watermark is embedded in the text as metadata and remains intact when text is copied and pasted elsewhere. Claude versions published after 2 August 2026 already include the feature, while older versions will be updated later. The marking is part of the text itself and does not change the meaning, quality, or readability of Claude's responses, according to the company. For certain file types, such as.png or.jpg images, Anthropic will also append digitally signed provenance metadata to indicate that the files were processed by Claude.

What the watermark does and does not prove

The watermark does not necessarily indicate that Claude authored a text entirely. Anthropic cautioned that users frequently employ Claude for proofreading, translating, summarising, or converting files, meaning a marked output could carry a Claude tag even when the underlying ideas, texts, or data originated from another source. The company warned against drawing hasty conclusions from detected markings. The watermark can survive some edits to a text, though it may not withstand heavy revision. Anthropic stated it would support users and third parties in detecting Claude's markings, as required by the EU code of conduct, with details to follow in future documentation.

EU AI Act compliance

The changes bring Anthropic into line with EU transparency rules that took effect on 2 August 2026. Brussels requires that AI-generated content carry machine-readable markings, with financial penalties for violations. A voluntary code of conduct has been signed by 190 companies, including Google, Meta, Microsoft, Mistral, and OpenAI. Anthropic is among the signatories. The company said all Claude models introduced in the EU from 2 August onward support machine-readable labeling from the start.

How the watermark may work

How the watermark functions technically has not been disclosed by Anthropic. One possible method, described by Bayerischer Rundfunk, is the "Green Flag/Red Flag" approach, which has been researched for several years. In this method, the language model splits its entire vocabulary into two halves at each word choice using a random key, giving one half a minimal statistical preference. No individual word reveals the pattern; only a longer text examined with the correct key shows the bias. The watermark may fail if the text is too short or has been heavily edited after generation.

Spotify and Google move in parallel

The audio streaming service Spotify also plans to improve AI labeling. From mid-September 2026, an "AI-Persona" label will indicate whether an artist profile represents real performers or AI-generated content. The label will appear on banners, in profile descriptions, in search results, and in playlists, following user concerns about profiles that appear human but are backed by AI-generated figures. Separately, Google announced that its AI chatbot Gemini has surpassed one billion monthly users. Google CEO Sundar Pichai posted on X on Tuesday:

This is our fastest-growing product of all time.