Does Claude watermark its text?

Yes, everywhere Claude is offered. Anthropic announced it on 14 August 2026: newer Claude models write text with an invisible watermark, and older ones are being brought in over the following months. Image files Claude creates get Content Credentials instead, a signed note in the file that anyone can read. Drop one in below.

✓ Free✓ No sign-up🔒 Deleted after readingReads Content Credentials

🔐

Drop or paste an image to read its Content Credentials

PNG and JPEG files from Claude carry a signed Content Credential. The reader shows the signer, and runs the AI detector when a file has none.

Choose an image →

This reads image files, not pasted text. Nobody outside the AI companies can check text for their watermarks yet.

Worldwide, not only in the EU

The EU AI Act asked AI companies to mark generated content from 2 August 2026. Anthropic applies the mark globally because, it says, it lacks a durable way to scope it by region. That covers the Claude apps, the API, Claude Code and Claude offered through AWS, Google Cloud and Microsoft Foundry.

Models launched in the EU from 2 August 2026 carry the watermark from launch. Models released before then are in a transition period. OpenAI took the opposite route and switched its watermark on for EU users only; see does ChatGPT watermark its text?

How the mark gets into the words

Claude normally uses randomness to pick each next word from the plausible options. With watermarking, that randomness comes from a key combined with the words already written, so the choices follow a pattern a reader cannot see but a detector with the key can. Anthropic says nothing is added to the text: no hidden characters and no extra tokens.

Text and files are marked differently

TextThe statistical watermark in word choice. Only detectable with Anthropic’s key. PNG, JPEG and SVG filesA signed C2PA Content Credential in the file’s metadata. The image itself is unchanged. Who can read itCredentials: anyone, with Anthropic’s own checker or the reader above (PNG and JPEG). The text watermark: approved organisations only.

Who can run the detector

Anthropic offers a detection API in private preview. Eligible organisations include regulators, law enforcement, media and fact-checkers, independent researchers, educational organisations and EU civil society groups, plus companies with their own compliance obligations. Others can register interest through a form on Anthropic’s announcement, and access is meant to widen over time.

Until then, a website that claims to spot the Claude watermark in pasted text is not reading it. Style-based AI text checkers guess; they do not hold the key.

When the mark is thin or missing

Short samples. Detection is unreliable on a few sentences and grows more confident with length.
Little freedom to choose words. Factual passages, code that must be exact and light proofreading of someone else’s writing change too few words to leave much of a pattern.
Editing. Light edits probably leave the mark in place. A full rewrite that replaces every word removes it.
Older models. Text from models in the transition period may not be marked yet.

What a mark says about authorship

A detected mark says Claude was probably involved. It cannot separate “Claude wrote this” from “Claude heavily edited this”, it does not say whether the rest came from a person or another AI, and it does not identify the user, the organisation or the conversation.

Some users objected when the change was announced, arguing that they wrote the instructions and Claude was only the tool. Anthropic’s position is that the mark changes neither ownership of the output nor responsibility for it.

Frequently asked questions

Does Claude watermark its text?

Yes. Since Anthropic’s announcement on 14 August 2026, newer Claude models add an invisible watermark to the text they write, worldwide. Older models are being added over the following months.

Can I check text for a Claude watermark?

Not as a member of the public, yet. Anthropic’s detection API is in private preview for eligible organisations such as regulators, media, fact-checkers, researchers and educators. Style-based AI text checkers cannot read the watermark.

Does editing remove the Claude watermark?

Light editing probably does not. A full rewrite that replaces every word does. Short passages, exact code and light proofreading carry too little of the pattern to detect reliably.

Does the Claude watermark identify me?

No. Anthropic says the watermark does not identify users, organisations or chats, and does not change who owns or is responsible for the output.

Do images and files from Claude have a watermark?

PNG, JPEG and SVG files Claude creates carry a signed Content Credential in their metadata rather than a watermark in the pixels. Upload a PNG or JPEG here to read it.