Anthropic adds hidden watermarks to Claude outputs

Anthropic is adding automatic watermarks to AI-generated text, images and files from Claude, tying the change to the European Union AI Act’s transparency code. The code took effect on Aug. 2 and is aimed at helping people distinguish human-made content from AI-generated material.

The company said supported Claude models weave an imperceptible watermark directly into generated text without changing its meaning, quality or readability. Anthropic said the markers can travel with copied and pasted text and may survive some editing, with the policy applying wherever Claude is offered, worldwide, including the Claude Platform API, Claude Code, Claude Cowork and Claude Tag.

Anthropic acknowledged that detection has limitations and may not always identify AI-generated content, particularly when writing has been heavily edited, paraphrased, translated, mixed with other writing or kept very short. Some users criticized the policy online, while others argued that open-source models without watermarks could offer easy workarounds.

The European Commission said it had strong tech industry backing for labeling AI-generated content to reduce deception and misinformation. As of July 31, the EU had secured commitments from nearly 200 companies, including OpenAI, Meta, Google and Microsoft.