Claude Text Watermark: How It Works, Who Can Check

Diagram of how a Claude watermark hides a statistical pattern in AI-written text

Nearly all of Claude’s output now has a concealed signal in it. The Claude watermark is a statistical pattern buried in word choices, and for now just the large AI labs and a handful of groups they’ve cleared are able to test for it.

Anthropic’s support page says its newest models and a few older ones mark their text on every product that offers Claude, in every country. The mechanism is simple enough that most developers could sketch it from scratch.

Why the labs are doing this

There are two stated reasons. One is practical: a machine-readable mark could let someone build filters that catch mass-produced AI text and make it harder to flood the web with it. The other is legal. The EU AI Act now requires AI providers to mark generated content in a form machines can read.

Anthropic isn’t alone. Google has marked Gemini app text since 2024. OpenAI said in October that ChatGPT and Codex text in the EU will get a mark too.

None of them lets the public run a detector on text. Google’s synthid.com checks images, video and audio, including pictures made with ChatGPT, but it won’t test text from any AI lab.

Claude watermark image 1
Image: oreillyradar.substack.com

The part of a model that matters

A language model predicts the next token. A token is a word or word fragment, and every model has a fixed vocabulary of them, each with a numeric ID. GPT-2, an early OpenAI model from 2019, has 50,257 tokens. Its tokenizer turns ” strawberry” with a leading space into a single token, number 41236, while “strawberry” with no space splits into three pieces: st, raw and berry. That is part of why models struggle to count letters. They never see letters, only numbers.

What the model hands back is a likelihood score for each token in its vocabulary, summing to 100%. It doesn’t choose a word. A separate program, the inference runtime, does the picking. It can take the likeliest token every time or roll weighted dice. All the randomness people notice in chatbot answers comes from that step.

This is where a watermark goes in. It doesn’t touch the model, only the runtime’s choice.

Be the watermark
Image: oreillyradar.substack.com

The green list method

The best-known early approach comes from a 2023 University of Maryland paper by John Kirchenbauer and colleagues. The runtime splits the vocabulary into a “green” and a “red” list, then boosts the probabilities of green tokens just before picking. Red tokens aren’t edited directly, but their share shrinks because the total must stay at 100%.

The split can’t be fixed, or readers could spot the favored words. Instead, a secret key combined with the previous token decides the green list at every position. To check a text, a detector rebuilds each list from the key and the preceding token and counts how many tokens landed on green. If the list is half the vocabulary, ordinary text would score about 50%. Watermarked text scores noticeably higher, and a statistical test shows how unlikely that is by chance.

The detector needs only the key and the tokenizer, not the model.

What the Claude watermark actually uses

Anthropic says the Claude watermark is built on SynthID-Text, which Google DeepMind described in a 2024 Nature paper. It also works in the runtime, but it doesn’t boost any probabilities. Anthropic’s own description says it swaps out only where the randomness behind word choice comes from.

In rough terms, the runtime draws several candidate words instead of one. It then assigns each candidate a coin flip determined by the secret key, the candidate and the few words before it. The runtime then emits one of the candidates that landed on heads. Because the flip is repeatable, anyone holding the key can redo it later.

When every candidate is the same word, there is nothing to choose, so the mark only appears at points where several words were genuinely plausible. A detector walks through the text, repeats each flip and counts heads. Ordinary writing gives heads about half the time. Watermarked writing gives more. Real detectors use more sophisticated statistics, but the idea is the same.

What this means for writers and readers

For now, nobody outside the approved circle can confirm whether a given paragraph came from Claude. Heavy human editing of an AI draft may well blur the pattern, though how much is unclear.

Detection by the public would change the picture. A public checker for text would let platforms, teachers and editors filter automatically, which is the use case that motivated watermarking in the first place.

What happens next

The gap is mainly one of access. The labs hold the keys. Whether Anthropic, OpenAI or Google open their text detectors to everyone is the open question. Until one does, the watermark exists but most people have no way to read the Claude watermark.

Source: oreillyradar.substack.com

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.