Claude starts invisibly watermarking everything it generates
Anthropic is embedding invisible watermarks in Claude-generated content to comply with the EU AI Act, rolled out worldwide. There is no self-check tool yet.
By Nattapon YongpaiboonCo-founder, Claude Thailand Community
What’s changing
Anthropic has started embedding a watermark in everything Claude generates, effective 2 August 2026. It applies across Claude.ai, Claude Code, Claude Cowork, the Claude Platform (API), Claude Tag, and when used through cloud partners AWS, Google Cloud, and Microsoft Foundry.
The main driver is Article 50 of the EU AI Act, which requires AI providers to mark generated content. Anthropic chose to apply this measure to everyone worldwide, not just users in Europe.
Two forms of marking
Text - an imperceptible pattern embedded in text Claude writes. It does not change the meaning, quality, or readability, and travels with the text when copied and pasted elsewhere.
Files - generated .svg, .png, and .jpg files carry signed metadata following the C2PA standard, an international standard for verifying digital file provenance, which can help detect later tampering.
Which models are covered
Models launched on or after 2 August 2026 carry the watermark from launch. Anthropic says it is retrofitting the capability into older models during the EU AI Act’s transition period, but has not given a firm timeline.
The key limitation: no way to self-check yet
Anthropic has not released a tool for users or third parties to detect the watermark themselves. It says only that it is working toward making detection possible and will publish technical documentation later.
Even once detection becomes available, some limits are worth keeping in mind:
- Finding a mark only means content may have been processed by Claude, not full proof of origin.
- Not finding a mark doesn’t mean content isn’t AI-generated. Heavy editing, file conversion, or unsupported platforms can strip the mark.
- Very short passages may not carry enough text for the watermark to be embedded reliably.
The author’s take
This matters for anyone using Claude to write client-facing work, or anything they’d rather not be known as AI-written. Even though there’s no way to check it yourself yet, it’s worth knowing the watermark has been there since 2 August 2026.
Details in this article come from Anthropic’s Help Center. Read the original source via the link below.
Get it by email
New articles, Claude updates and community event announcements. Sent occasionally, never often enough to annoy you.
The newsletter is written in Thai