May 14, 2024 - Technology

Google devises new watermark for AI-made text

Illustration of a rubber stamp next to an inked rectangle with the letters AI inside

Illustration: Natalie Peeples/Axios

Google Tuesday unveiled a new technique for marking text so it can later be identified as AI-generated without altering the quality or meaning of the writing.

Why it matters: As AI-created content begins to flood the internet, the need to detect and distinguish it has grown — but watermarking tools have been much easier to develop for images than for text.

The big picture: The new features for Google's existing watermarking tool, called SynthID — which until now has only handled images and audio — will automatically watermark text generated by Google's Gemini AI programs in the Gemini app or on the web.

  • Google says in its announcement that the tool will work best with prompts for longer and more varied text, like writing an essay, a theater script or "variations on an email."
  • SynthID will also watermark video generated by Veo, the company's generative video model.

Catch up quick: Combating misinformation, copyright violations and plagiarism requires tools that can detect AI-generated images, video and text.

  • A popular technique to avoid AI misuse is to add invisible watermarks to the AI-generated content, and SynthID, launched last year, was one tool to create them.
  • For images, Google designed the watermark to remain detectable even after they were modified by adding filters, changing colors or adjusting brightness. And unlike visible watermarks, SynthID can't be removed by cropping.

How it works: Large language models for text generation work by producing sequences of text in response to prompts.

  • LLMs predict the next phrase, word, or character based on which is the most likely to appear next in the sequence.
  • The phrases, words and characters are called "tokens" and each token is given a probability score. "The final pattern of scores for both the model's word choices combined with the adjusted probability scores are considered the watermark," Google says.
  • "It's specifically changing the AI generated content in a way so that it remains detectable in the future," Pushmeet Kohli, VP or research at DeepMind, told Axios.

Yes, but: Google says AI-generated text with its watermark can be detected even with "mild paraphrasing," but the watermark is less detectable when content has been thoroughly rewritten or translated.

  • SynthID for text is also less effective for factual material when there are limited potential responses to a given prompt.

What we're watching: Later this summer Google plans to open source this watermarking technique so that others can incorporate it into their services.

  • Although Google says SynthID for text is "compatible with most AI text generation models," it remains to be seen if the company's competitors — including ChatGPT, Microsoft and Meta — will adopt it or come up with their own approaches.
Go deeper