Anthropic has announced that its new models will insert machine-readable marks into content produced by Claude in order to comply with European Union transparency rules. The company said these measures apply to models launched in the EU after Aug. 2, 2024, and will be implemented “wherever Claude is offered, worldwide.”
What the markings are
Anthropic described two complementary approaches:
- Text watermarks: Claude embeds patterns into generated text that Anthropic says are imperceptible to human readers but detectable by automated tools.
- File metadata signatures: generated media files carry digital signatures indicating the asset was processed by Claude.
Why it matters
The changes are intended to meet Article 50 of the EU AI Act, which requires disclosures when content is generated or manipulated by AI. The new markings could affect ordinary editorial workflows: communications teams that use Claude to proofread, translate or reformat human-authored press releases might still produce documents that carry an AI signature.
Limitations noted by Anthropic
Anthropic warned of two main limitations in its detection approach:
- AI-assisted but human-origin content can look AI-generated: content may receive a detected mark even if Claude was used only to proofread, format or translate text originally written by a human.
- Detection falloff: if generated text is heavily rewritten, combined with other material, or is too short, the watermarks may become undetectable.
Broader context and industry moves
Anthropic framed this change as a compliance-driven response to the EU law. OpenAI has also outlined its approach to compliance, though its current watermarking work emphasizes images and audio more than text.
Platforms have accelerated their own AI-identification measures: LinkedIn is testing a "seems like AI slop" button, Substack integrated Pangram’s AI-detection suite, and Snap has stopped promoting AI-generated video in its main feed.
What’s next
AI providers have until Dec. 2, 2024 to bring legacy models into compliance with the EU rules. In the coming months, the industry’s implementation choices should become clearer as companies update older models and report how they mark AI-generated or AI-modified content.
Overall, Anthropic’s announcement represents a regulatory-driven shift toward more explicit identification of AI-produced text, audio and visual material, while acknowledging technical limits to reliably detecting such content.
