Anthropic will embed invisible digital watermarks in text and images produced by Claude to comply with the EU AI Act. These identifiers will enable automatic detection of AI-generated or modified content, with phased implementation by December 2026.
A legal requirement from Europe
The EU AI Act requires AI service providers to make content generated or modified by their tools identifiable. The deadline for new model versions was August 2, 2026. Anthropic, like OpenAI and Google before it, has published its guidelines to comply with Article 50 of the law and the accompanying Code of Practice.
The watermarking will apply to text and images (SVG, PNG, and JPG files). For text, the method relies on a statistical bias applied during token generation (word fragments). The result follows a detectable pattern via dedicated tools, without affecting content quality or meaning. Excerpts under 200 tokens (about 150 words) are exempt, as they are too short for reliable watermarking.
Images: provenance certificate, but no digital watermark yet
For images, Anthropic will add a provenance certificate in C2PA format, a standard that verifies the file’s origin. This certificate becomes invalid once the image is modified. However, the Code of Practice also requires a digital watermark embedded directly into the visual content — a measure the company has not yet announced but should introduce to prevent images from becoming untraceable after a simple copy.
Global rollout and technical limitations
Anthropic states that watermarking will first apply to models launched in the EU but will later extend to all versions of Claude, regardless of where they are used. Existing models must be updated by December 2, 2026. The company acknowledges that direct quotes might be incorrectly watermarked and that the absence of a watermark does not guarantee content was not AI-generated or edited.
Reactions: support and resistance
On Reddit, some users criticize the measure, calling it “draconian” or “hypocritical” given the data used to train the models. Others see it as necessary protection against deceptive uses, such as pasting AI-generated text into schoolwork or articles. A minority argues that watermarking could harm legitimate uses, like writing assistance or summarization.
“Who will get caught? You. The student who used Claude to reorganize a paragraph. The journalist who asked AI to summarize a 200-page report. The writer who had a creative block and asked for synonyms. These people come out of the process with a digital watermark on their forehead.” — visionode, Reddit userTranslated from French
Watermarking AI content is becoming standard, but its effectiveness will depend on its resilience to modifications and the adoption of detection tools. Anthropic promises technical details in upcoming documentation.
Comments 0
Sign in to join the discussion.
Be the first to comment.