AI-generated text now carries an invisible watermark. What is it good for?

AI companies must now add watermarks to text and images generated by their models under a European requirement introduced last month. The aim is to counter misinformation and manipulation, but the watermark remains invisible to readers.

Anthropic, the US company behind the Claude chatbot, has explained its approach in a detailed blog post. The technology turns out to be a version of SynthID, a system developed by Google DeepMind researchers in 2024. Google’s Gemini chatbot has added a watermark since November, while OpenAI says it is working on a similar system for ChatGPT text.

The watermark lies in the model’s choice of words. A large language model predicts what word should follow the words already generated, taking the prompt and many other variables into account. Often, several words or phrasings would work equally well. A chatbot can therefore produce different suitable answers to the same question.

SynthID uses that built-in randomness. If “cloudy” and “grey” both fit the sentence “the sky is...”, the model assigns each option a random score and chooses according to probability. Without a watermark, the scoring is entirely random. With SynthID, a secret key slightly influences the scores, giving one option a small statistical advantage. The preferred word does not become mandatory, so the text remains varied.

“This is really cleverly done, because this watermarking method mathematically ensures that the quality of the output does not decline,” said Rik van Noord, an AI researcher and large-language-model expert at the University of Groningen. All selected words remain within the model’s list of options considered suitable for the context, he said. Google DeepMind researchers reached the same conclusion after analyzing 20 million texts generated by Gemini and comparing them with other material.

“

This is really cleverly done, because this watermarking method mathematically ensures that the quality of the output does not decline.