Since August 2, 2026, new Claude models mark what they generate: an imperceptible watermark inside text, and C2PA signed provenance inside generated files. Here is what can actually be verified today — and what can't yet.
What Anthropic actually says
Under the EU AI Act’s transparency code, Anthropic documents that supported Claude models weave an imperceptible watermark directly into generated text, and that files Claude generates in supported formats (SVG, PNG, JPG) carry signed C2PA provenance metadata. Anthropic states it will enable users and third parties to detect these marks, with technical documentation to follow.
Official detection is the next step: Anthropic has committed to letting users and third parties verify Claude’s marks, with technical documentation on the way.This checker is built around that mechanism and will run it the day it’s released — same URL, one more check. Until then, any tool claiming to “detect the Claude watermark” in text is guessing. What you can genuinely verify right now:
Claude-generated files — if Claude produced a PNG, JPG or SVG for you, its C2PA Content Credential can be validated in the image checker below.
Hidden characters in text — Claude’s watermark is statistical, not character-based, but the text inspector below still reveals any invisible Unicode markers from other sources.
Check a Claude-generated file
Inspect text for hidden characters
Show the hidden characters inside the text
Copy cleaned text
Removes only the characters you select below (unusual spaces are replaced with a normal space). This is a transparency tool — it does not remove statistical watermarks and is not a way to hide AI use.
Official provider verificationClaude detection: announced
ClaudeWatermark live in models since Aug 2, 2026. Public detection is announced — it plugs in right here the day Anthropic ships it.
This page runs each official check the moment it becomes available — follow the rollout on the provider status page.
What a detected mark would mean
Anthropic is careful about this, and so are we: a detected Claude mark is evidence the content may have been processed by Claude — not that Claude authored every idea in it. Content can be edited after marking; heavy editing or translation can make marks undetectable; and the absence of a mark never proves text wasn’t AI-generated.
When Anthropic ships public detection
This page will integrate the official mechanism as soon as it is technically and contractually possible, clearly labeled as an official check. The URL will not change. Meanwhile the full picture across providers lives on the AI watermark status page, and the background is explained in Does Claude watermark text?