ClaudeAI watermarking

Aug 14, 2026 · Tech/AI · 2 min read · #ai-content-detection

How Claude marks AI generated content

Anthropic signing the EU AI Act’s Article 50(2) Code of Practice on transparency of AI-generated content, and what that means for Claude.
Claude models launched from 2 August 2026 supporting marks from day one, across the API, Claude, Claude Code, Claude Cowork and Claude Tag, worldwide, and on AWS, Google Cloud and Microsoft Foundry.
A table of which Claude models support text watermarks and Content Credentials, with older models being added.
Two techniques: an invisible watermark woven into generated text, and C2PA Content Credentials on generated files such as PNG and JPEG.
Detection through the free Claude Content Checker for files, and watermark detection in private preview for regulators, media, researchers and similar groups.
Limits: a mark doesn’t prove Claude wrote something, and no mark doesn’t mean content wasn’t AI-made, plus a note for those building with Claude.

Kuditipudi, Thickstun, Hashimoto and Liang’s method for watermarking language model text that survives edits without changing the output distribution, up to a set generation budget.
Mapping a sequence of random numbers from a secret key to model samples, so anyone with the key can detect the watermark by aligning the text to that sequence.
Two sampling schemes: inverse transform sampling and exponential minimum sampling.
Tests on OPT-1.3B, LLaMA-7B and Alpaca-7B, detecting watermarks from 35 tokens even with 40-50% of tokens randomly edited.
The Alpaca-7B case study, where lower-entropy instruction responses make detection harder and less robust to paraphrasing.