What Is Text Watermarking? How AI Text Gets Marked and Checked

LLM foundationsSafety and governancePublished By Simon Budziak

Text watermarking hides a statistical signal in AI-generated text so software can later check which model wrote it. The signal sits in word choices rather than visible marks, so it is useful for provenance but weaker on text that people have edited.

It is a narrower idea than AI watermarking, which also covers images, audio and video. A text watermark is not the “DRAFT” stamp you place across a Word page, and it does not read as a mark at all.

How does a text watermark work?

The generating model nudges its own word choices. In the method of Kirchenbauer and colleagues, a random “green” subset of the vocabulary is chosen before each word, and the model gently raises the odds of green words; a detector then counts green words and applies a statistical test that yields a p-value (arXiv 2301.10226, read 19 November 2026). Detection needs no access to the model, only the key that defines the green lists. The more text, the stronger the evidence.

Anthropic says text from supported Claude models carries its own text watermark, which adds no tokens or hidden characters (Anthropic docs, read 19 November 2026). Google DeepMind’s SynthID-Text adjusts token probability scores in a similar spirit. It was released as open source and integrated into Gemini in October 2024; the signal can be found in as few as three sentences, though longer text is more reliable (The Decoder, 23 October 2024, read 19 November 2026).

Where does text watermarking fail?

A watermark survives copying but weakens with editing. The same report says detection is weaker on edited text. Short snippets carry too little signal to test. A watermark also proves nothing about text that was never watermarked: a failed check does not mean a human wrote it.

Check What a positive result tells you What a negative result tells you
Watermark detector This text probably came from the marking model Nothing reliable
AI content detector A classifier guesses the style looks machine-made Nothing reliable
C2PA provenance record A signed history of how the file was made The history was stripped or never added

What should a business ask about text watermarking?

Treat a watermark as one signal in a provenance process, never as proof. Ask a vendor which models and languages it covers, how long a text must be for detection, who holds the key, and what happens after edits. For documents you publish, pair it with content provenance records and a rule on who may use generated text, set in your AI usage policy. If a decision, such as disciplining an employee or rejecting a submission, rests on the result, a human has to weigh other evidence first.

This entry was drafted with AI assistance.

Frequently asked questions

Does ChatGPT leave watermarks in text?

Do not assume it does. The best-known deployed text watermark is Google's SynthID-Text in Gemini (The Decoder, 23 October 2024, read 19 November 2026). Anthropic also watermarks text from supported Claude models ([Anthropic docs](https://platform.claude.com/docs/en/build-with-claude/text-watermark), read 19 November 2026). For any other tool, ask the vendor in writing whether it marks text and how you would check.

Can a text watermark prove a person used AI?

No. A detector reports how likely a marked model wrote the text, and edits or short length can weaken the signal, so a missing watermark proves nothing either way.

Summarize this page with

Train your team to build this