Google tools make AI-generated posts easily detectable

Google uses artificial intelligence watermarks to automatically identify text generated by its Gemini chatbot, making it easier to distinguish between AI-generated content and human-written posts. This watermarking system could help prevent AI chatbots from being exploited for misinformation and disinformation, as well as fraud in schools and business environments.

Pushmeet Kohli of Google DeepMind, the company’s AI research team, said the company is now making available an open source version of its technology, and that other generative AI developers can do the same by drawing from their large-scale language models. It is possible to add watermarks to the output. It was once a research lab for Google Brain and DeepMind. “SynthID is not a silver bullet for identifying AI-generated content, but it is an important building block for developing more reliable AI identification tools,” he says.

Independent researchers expressed similar optimism. “No known watermarking method is foolproof, but I really think this will help catch some of the misinformation that AI generates, academic misconduct, etc.” said Scott Aaronson of the University of Texas at Austin, who has studied safety. . “We hope that other leading language modeling companies, such as OpenAI and Anthropic, will follow DeepMind’s lead in this regard.”

In May of this year, Google DeepMind announced that it had implemented a SynthID method for watermarking AI-generated text and video from Google’s Gemini and Veo AI services, respectively. The company recently published a paper in the journal nature SynthID generally performs better than similar AI watermarking techniques on text. The comparison involved evaluating how easily the responses from different watermarked AI models were detectable.

In Google DeepMind’s AI watermarking approach, as a model generates a sequence of text, a “tournament sampling” algorithm subtly moves it toward selecting “tokens” of specific words that are detectable by associated software. Create a statistical signature. This process randomly pairs candidate word tokens in tournament-style brackets. The winner of each pair is determined by which one gets the highest score according to the watermark function. Winners advance through successive tournament rounds until there is one round remaining. It’s a “multi-layered approach,” and “the possibility of trying to reverse engineer or remove the watermark adds complexity,” said Flon Huang of the University of Maryland.

Harvard University’s Hanlin Zhang said a “determined adversary” with vast computational power could remove such AI watermarks. But he said SynthID’s approach makes sense given the need for scalable watermarking in AI services.

Google DeepMind researchers tested two versions of SynthID that represent a trade-off between making watermark signatures easier to detect in exchange for distorting the text typically produced by AI models. They showed that the undistorted version of the AI ​​watermark continued to work without noticeable impact on the quality of the 20 million text responses Gemini generated during live experiments.

However, the researchers also acknowledged that this watermarking works best on long chatbot responses that can be answered in a variety of ways, such as composing an essay or an email, as well as on math or coding questions. The response to this has not yet been tested.

Google DeepMind’s team and others have stated the need for additional safeguards against misuse of AI chatbots, and Huang similarly recommended stronger regulation. “Requiring watermarks by law addresses both practicality and user adoption challenges and makes large language models more secure to use,” she says.