Text Watermarking for Non-Academics

Text Watermarking for Non-Academics

Wed Aug 12 2026

tutorials

artificial intelligence

language models

text watermarking

provenance

AI detection

Anthropic has just announced how Claude marks AI-generated content, bringing text watermarking into immediate practical discussion. A watermark in plain text is less intuitive than one in an image or a file because copied text appears to leave its provenance behind. These two articles examine the mechanism and its implications. This first article explains how statistical text watermarking works from first principles, the second examines what detecting such a watermark would allow us to conclude.

How Do You Watermark Plain Text?

Images provide an obvious place for a watermark: a system can alter pixel values by amounts too small for a person to notice, while a detector can still recover the resulting pattern. Digital files can also carry metadata that records their origin or editing history. Plain text appears to offer neither option because it consists only of characters on a page.

Copy a paragraph from a document into Notepad and save it as a new file. The operation discards most of the surrounding information, including document metadata and details of the original formatting. What remains is the sequence of words and punctuation that the reader can see. Any marker that depends on the original file has disappeared.

The text still contains more information than its literal meaning, however. A writer repeatedly chooses between words such as but and however, contractions such as don't and do not, and several ways of arranging the same thought. Sentence length, punctuation, vocabulary, and grammatical construction introduce further choices. Each version can communicate substantially the same idea while leaving a different sequence of characters behind. Natural language as an encoding system is inherently redundant, and that redundancy allows a writer to leave a statistical signal in the text itself .

A language model makes comparable choices whenever it generates text. Given the passage produced so far, the model assigns probabilities to possible next tokens and selects one according to its generation settings. A token can be a whole word, part of a word, a punctuation mark, or another text fragment. The visible sentence is therefore the result of many selections from many possible continuations.

A simple completion illustrates the available freedom. The list below contains several plausible continuations for the same opening:

The result was ...

unexpected<br>surprising<br>interesting<br>quite different<br>hard to explain<br>...

Several completions can be fluent and appropriate. Selecting one of them changes the text without requiring hidden characters or attached metadata. The chosen words themselves preserve the selection after someone copies the passage into another document.

Text watermarking uses this freedom in a controlled way. During generation, a watermarking system can influence which acceptable continuation the model selects. Across a long passage, those influenced choices form a statistical pattern that a detector can test for later. The pattern lives in how the text was chosen, so it remains part of the text after ordinary copying and pasting.

An individual choice reveals almost nothing. A person can write however, and an unmodified language model can select it as well. Evidence starts accumulating when many small choices lean in a coordinated direction more often than chance would predict. This dependence on accumulated evidence is why statistical text watermarking belongs to a different category from a visible label or a fixed hidden message.

Anthropic's announcement gives the subject new practical relevance, but the general technique is broader than Claude or any other product. Vendors can use different token-selection rules and detection methods, and public descriptions do not provide a complete implementation specification. The principles in this article describe the technology rather than attributing a particular design to Anthropic.

Humans Already Have Statistical Fingerprints

Of course the choices encoded in a piece of text existed long before language models. Every writer develops habits through education, reading, geography, profession, and personal preference. Some habits are easy to notice, while others appear only after someone counts them across a substantial body of work. Apparently, a text without any pattern would be very difficult to read. Two writers expressing similar ideas can consistently favour different forms. Their tendencies might look like this:

Writer A Writer B

but however<br>for example for instance<br>don't do not<br>shorter sentences longer sentences<br>few semicolons frequent semicolons<br>simple punctuation lots of em dashes (the classical LLM watermark)

Finding however in a document provides no useful attribution on its own. Writers use words outside their usual habits, and the subject...