Just published in @Nature! Research by @GoogleDeepMind including Nobel Laureate @demishassabis introduces a novel technique for reliably identifying LLM-generated content. At its core, this approach subtly alters the next-token sampling process to introduce minor adjustments into the distribution of generated text. These slight modifications embed a “signature” in the LLM-generated text that would be nearly invisible to a human reader but detectable using specific statistical tools. A major advantage of this method is its efficiency: detection doesn’t require resource-intensive computations or direct access to the underlying LLM, which is often proprietary. Full paper: nature.com/articles/s41586-0…