IsWatermarked?

Does Claude Watermark Text? What Anthropic's 2026 Marking Actually Means

By IsWatermarked Research · August 18, 2026

Abstract illustration of ink ripples spreading across paper, suggesting an invisible mark inside text
Illustration: IsWatermarked? original

Short answer: yes. Anthropic documents that supported Claude models launched on or after August 2, 2026 mark the content they generate — and that includes an imperceptible watermark embedded directly in generated text. Models released earlier are being brought up to the same behavior. Here is what Anthropic actually says, what it means in practice, and what nobody can check yet.

What Anthropic deployed

Anthropic committed to the EU AI Act’s transparency code of practice, and its help-center documentation describes two distinct marking mechanisms:

  1. An embedded text watermark. This is a statistical watermark — a pattern in how words are chosen — not hidden characters sprinkled into the output.

“It weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning.”

— Anthropic Help Center · How Claude marks AI-generated content · 2026

  1. Signed file provenance. When Claude generates files in supported formats (SVG, PNG, JPG), it attaches signed provenance metadata following the C2PA standard. This is the same Content Credentials system used by OpenAI, Adobe and camera manufacturers.

The marking applies across supported Claude surfaces — the consumer apps as well as generated artifacts.

Can you check for the Claude watermark today?

Here is exactly where verification stands:

  • The text watermark: detection is coming. Anthropic states it will “enable users and other third parties to detect Claude’s embedded watermarks,” with technical documentation to follow. The Claude watermark checker on this site is built to run that official mechanism the day it ships. Until then, anything marketed as a “Claude watermark detector” for text is either inspecting hidden characters (a different thing) or running a statistical AI-detector guess (also a different thing — here’s the distinction).

  • Generated files: yes. If Claude produced a PNG, JPG or SVG for you, its C2PA credential is verifiable right now, locally, in the image checker. A valid signature proves the provenance record is authentic and the file hasn’t been modified since signing.

  • Hidden characters: checkable, but a separate question. Claude’s watermark is not made of zero-width characters. Still, text that has traveled through chats, docs and websites accumulates invisible Unicode — the text checker makes all of it visible.

What a detected mark will (and won’t) prove

Anthropic is unusually careful here, and the caveats are worth quoting in substance:

  • A detected mark is evidence the content may have been processed by Claude — not that Claude originated every idea or sentence in it. A human can heavily edit Claude output; a human can also paste their own writing through Claude.
  • Marks are designed to survive ordinary handling, but heavy editing or translation can make them undetectable.
  • Absence of a mark proves nothing. Text could come from an older model, a different provider, or have been transformed enough to strip the pattern.

In other words: even after official detection ships, the honest output of a checker is evidence with context — never a clean “written by AI” / “written by a human” verdict.

Why this matters more than previous “AI detection”

Classifier-based AI detectors read style and guess; providers’ embedded watermarks are designed-in signals that the provider can verify cryptographically or statistically with known error bounds. A working public Claude detector would be a categorically more reliable signal than any AI-detection startup has ever offered — which is exactly why we track its release on the Claude watermark checker page and integrate it the moment supported third-party detection exists.

This article states what is verifiable from primary sources as of its last-updated date above. Provider behavior changes; the status page is maintained accordingly.