How does Anthropic watermark images and files?

For images and other files, Anthropic uses the C2PA standard for digitally signed provenance metadata, rather than a text watermark. C2PA cryptographically records that content was AI-generated and by which system.

Why is Anthropic adding watermarks?

It is a compliance move under the EU AI Act's Article 50 transparency rules, which require AI-generated content to be machine-detectable. The obligation targets new Claude models launched in the EU from August 2, 2026 onward.

Quick Answer

Anthropic to Watermark Claude AI Text (Aug 2026)

Q: Is Anthropic watermarking Claude's output?

Yes. As of August 2026, Anthropic will embed invisible, machine-readable watermarks in text generated by Claude, driven by the EU AI Act's Article 50(2) Code of Practice on transparency of AI-generated content. It applies to new Claude models launched in the EU on or after August 2, 2026, with work underway to extend it to older models.

Published: August 12, 2026

The Short Answer

As of August 2026, Anthropic will embed invisible, machine-readable watermarks in text generated by Claude, driven by the EU AI Act’s Article 50(2) Code of Practice on transparency of AI-generated content. It applies to new Claude models launched in the EU on or after August 2, 2026, with C2PA provenance for images and files.

What Changed

The EU AI Act requires AI-generated content to be machine-detectable. Anthropic’s answer: invisible text watermarks for Claude’s writing and the C2PA standard (digitally signed provenance metadata) for images and other files — so downstream tools can verify content came from an AI system.

Key Facts (as of August 2026)

	Anthropic watermarking
Driver	EU AI Act, Article 50(2)
Text	Invisible, machine-readable watermarks
Images/files	C2PA signed provenance metadata
Applies to	New Claude EU models from Aug 2, 2026
Older models	Extension work underway

Why It Matters

This is the EU AI Act biting in practice — not a research demo. Watermarking becomes a shipping requirement for frontier vendors selling into Europe. Expect the pattern to spread: invisible text signals plus C2PA for media are becoming the de facto compliance stack for AI provenance.

The Reality Check

Invisible text watermarks are fragile — heavy editing, paraphrasing, or translation can degrade them. C2PA metadata can be stripped. The regulation forces a best-effort provenance signal, not a tamper-proof one. Still, it’s a meaningful shift: the biggest labs now treat detectability as a legal obligation, starting with the EU.

Sources

Trending Topics — Anthropic plans watermarks for AI content from Aug 2026: trendingtopics.eu
Fast Company — Anthropic models will inject watermarks in AI text: fastcompany.com
Newscord — Anthropic watermarks Claude text for EU AI Act Article 50: newscord.org