Researchers at Google DeepMind in London have developed an innovative 'watermark' to invisibly label text produced by artificial intelligence (AI), which is now in use by millions of chatbot users. This watermark, detailed in a Nature publication on October 23, represents a significant real-world application, despite not being the first of its kind or immune to removal attempts. According to Scott Aaronson, a computer scientist formerly with OpenAI who worked on AI watermarks, the deployment of this technology is noteworthy. The ability to identify AI-generated text is increasingly critical for addressing issues like fake news and academic dishonesty, as well as preventing the degradation of future models through training on AI-generated content. In extensive trials involving Google’s Gemini large language model (LLM), users indicated that they considered watermarked texts to be of similar quality to those without watermarks. Experts suggest that watermarked tools are likely to proliferate in the commercial space. DeepMind’s watermark, SynthID-Text, is unique in its approach, as it modifies word selection in a systematic but hidden manner that can be detected using a cryptographic key.

This method is more detectable than competitors' solutions while ensuring that text generation remains efficient. Pushmeet Kohli of DeepMind encourages other AI developers to incorporate this watermarking system into their models, although Google maintains confidentiality regarding the detection key. While governments view watermarking as a means to combat AI-generated text proliferation, challenges persist in achieving uniform adoption across developers and addressing vulnerabilities. Research has indicated that all watermarks risk being removed or imitated. DeepMind's watermarking technique employs a sampling algorithm that builds on existing methodologies, wherein a series of token comparisons, likened to a tournament, determines which token to generate. This complex process enhances watermark detectability and complicates removal attempts; researchers demonstrated that even after text paraphrasing by another model, identification of the watermark remained viable. However, the effectiveness of this watermark against intentional removal has not been thoroughly examined, presenting a significant unanswered question in AI safety. Kohli emphasizes the intention behind the watermark as a tool to facilitate responsible LLM use and foster community-driven improvements.