Anthropic says its newest Claude models and some older versions watermark generated text worldwide, across every surface where Claude is available. The machine-detectable signal could help filter mass-produced AI content, but access to the detector needed to check it is currently limited to major AI companies and a handful of organizations they have approved.
Andrew Stellman, the source article's author, points to filtering the flood of low-quality AI writing online as one possible use for watermarking. Anthropic and OpenAI also cite the European Union's AI Act, which requires AI-generated content to be marked in a machine-readable way.
Google has used a text watermark in its Gemini app since 2024. OpenAI announced in October that it was adding one to text produced by ChatGPT and Codex in the European Union. None of the companies makes its text detector available to the public. Google's SynthID service allows anyone to check images, video and audio, including images made with ChatGPT, but not text.
Tokens and the next-word choice
Understanding watermarking requires looking at the output stage of a language model. Models do not work directly with letters: a tokenizer divides text into words or word fragments and assigns each a numerical identifier.
OpenAI's 2019 GPT-2 model, for example, has a vocabulary of 50,257 tokens. The word strawberry with a leading space is a single token with the ID 41236; without the space, its tokenizer splits it into st, raw and berry. This processing can contribute to models' difficulty counting letters.
Token IDs go into the context window, which in a conversation generally contains previous messages and answers as well as instructions or files added by the application. Based on that context, the model assigns a probability to every possible next token. A separate inference runtime makes the actual selection: it can choose the most likely token or sample randomly according to the probabilities. The selected token is then added to the context and the process repeats.
Two approaches to watermarking
In a 2023 paper, John Kirchenbauer and colleagues at the University of Maryland described a method known as the green list. Before selecting the next token, the runtime increases the probabilities of a group of tokens determined using a secret key. The favored group changes according to the key and the preceding token, so the same words are not always favored.
A detector can use the key and tokenizer to reconstruct the lists for each position. If an unusually high number of tokens in a passage are on their respective green lists, that can statistically indicate a watermark; the model itself is not needed for the check.
Anthropic says Claude uses a different approach: a version of SynthID-Text, which Google DeepMind described in the journal Nature in 2024 and also uses in the Gemini app. Anthropic says its watermark changes the source of randomness used to choose among words.
In a simplified explanation, the runtime samples several candidates for the next token and assigns each a repeatable “coin flip” based on a secret key, the candidate and preceding tokens. It favors a candidate whose flip comes up heads. If all the candidates are identical, there is no real choice, and the signal cannot be introduced at that point.
A detector with the key can later repeat the calculations. Ordinary text would be expected to produce heads at roughly half the positions, while watermarked text produces heads more often. Real detectors perform a more sophisticated analysis than this illustration suggests.
A simplified experiment
Stellman also illustrates the principle with an interactive web page that can be built using an AI assistant. In the example, many words in a short story can be swapped for alternatives with similar meanings. The toy checks each word using a five-digit key, 3 1 4 1 5: it multiplies the IDs of the word and the four preceding words by the key's digits, then adds the results. If the last digit of the sum is between 5 and 9, the imaginary coin comes up heads.
The starting story produces heads at 14 of 32 checked positions, close to the half expected by chance. A user can increase the proportion by swapping words; the toy reports a watermark at 60% or more, which means 20 heads. A swap also affects the next four calculations, so working from left to right helps. The first four words are not checked because there are not enough preceding words for the calculation.
Changing the last key digit to 6 will likely bring the heads count closer to half and may make the indication disappear, even though the pattern made with the original key remains. A positive result is still possible by chance with the wrong key.
The experiment also shows why longer passages make for more reliable checks. With 32 independent, fair coin flips, at least 60% will come up heads by chance roughly once in nine trials; across hundreds of positions, the chance of that proportion is much smaller.
The toy is considerably simpler than real SynthID. Actual token IDs are not assigned alphabetically; the system uses a much larger key, combines it with the words through a hash function rather than a sum, and performs several rounds per token. Detectors weight those rounds differently and are trained for each key on both watermarked and ordinary text.
Why detection remains restricted
A secret key could be used not only to identify a pattern but also to weaken or remove it by editing the text, or deliberately introduce it into human-written text. A public key could therefore enable tools that automatically erase the marking.
That may explain why companies restrict access, although the restriction also impedes independent checks. Stellman writes that he is applying for access to Claude's detector.



