On 14 August 2026, Anthropic announced that future Claude models will watermark their text. Google's Gemini has carried a text watermark since 2024. Neither adds a character to the page. The watermark lives in which words the model chooses.
The method behind both is SynthID-Text, which Google DeepMind published in the journal Nature on 23 October 2024. This note walks through it one word at a time.
A model writes by drawing lots
A language model writes one word at a time. For each slot it works out a probability for every word it knows, then draws one. After "The garden was quiet and", the words "still", "calm" and "peaceful" all read fine, so the draw is settled by a random number. A page of chatbot writing is hundreds of these draws.
The paper's authors put the problem plainly: "it can become difficult to distinguish synthetically generated text from human-written text". A watermark lets a checker calculate whether the model would have drawn these words using a secret key held by the model's maker.
The watermark rigs the draw
SynthID-Text replaces the random number. Training, the part where the model learns to write, stays the same. Only the last step, the draw, changes. For each new word:
Make a seed. Combine the four words just written and the secret key into one number, called the seed. This calculation is called hashing. The same four words and the same key give the same seed every time.
Score every word. The seed sets a series of scoring rules. Each rule stamps every word in the vocabulary with a 0 or a 1. To anyone without the key, the stamps look like coin tosses.
Run a tournament. Draw eight candidate words from the model's own probabilities. Pair them off. In each pair, the word stamped 1 by the first rule beats the word stamped 0; a tie is settled by chance. The four winners pair off again under the second rule. The two winners meet under the third. The last word standing is written.
The paper illustrates the method with eight candidates and three rounds. DeepMind's experiments used thirty rounds. A word the model would never use cannot enter, because every candidate comes from the model's own list.
Checking a text needs the key, not the model
The checker reads the text one word at a time. For each word, the checker hashes the four preceding words with the key. It uses the result to rebuild the scoring rules and look up the stamp for the word being checked. Then it averages the stamps.
In text nobody watermarked, the stamps average about a half, because without the key they are coin tosses. In text picked by the tournament, the stamps average higher, because the tournament kept choosing 1s. The longer the text, the clearer the gap. The checker can give a confident answer for a long essay. For a two-line reply, it has almost no confidence in its answer.
Nobody noticed
DeepMind tested the watermark on Gemini itself. Some users' requests went to a watermarked model and others to the plain one, and the team compared the thumbs-up and thumbs-down clicks on about 20 million replies.
| Check | With watermark |
|---|---|
| Thumbs-up rate, about 20 million Gemini replies | +0.01% |
| Thumbs-down rate, the same replies | −0.02% |
| 3,000 answers from a small model anyone can download, rated side by side | No preference |
| Time per word on the same small model, thirty rounds | +0.57% |
With the setting tested here, the watermark leaves each word's chance of being chosen unchanged within one reply. Gemini uses this setting.
Where it fades
The watermark needs choices. Ask for the capital of France and there is one right answer, so the tournament has nothing to pick. Code is the same: most of it must be exact. DeepMind says the check works best on long and varied responses, and that its confidence can drop sharply when a text has been thoroughly rewritten or translated.
The check answers one question: how likely is it that this model wrote this text? The check cannot tell you whether a person wrote the text. It cannot recognise another company's model, which uses its own key.
Gemini has it. Claude is next.
DeepMind released the code under an open licence, and it ships in the Hugging Face Transformers library from version 4.46. Since 2 August 2026 the European Union has required AI providers on its market to mark generated content. Anthropic's announcement gives the EU AI Act as the reason and calls its method "a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024". Anthropic says the watermark carries nothing about the user or the chat and adds no cost. Anthropic's detection service is open, for now, only to regulators, researchers and others named in the EU law.