Watermarking LLM-Generated Text

A Watermark for Large Language Models studies selecting pseudorandom green vocabulary subsets and promoting their tokens during generation. The soft scheme adds a bias to green-token logits rather than requiring every sampled token to be green. A detector reproduces the relevant partitions and measures an excess of green tokens under a stated null hypothesis. Its implementation needs the matching scheme configuration, tokenization, and any watermark key used; access to the generation model’s parameters or original prompt is not inherently required for this detector. A simplified statistic subtracts the expected green count from the observed count and divides by the null standard deviation. With one hundred eligible scored tokens, green fraction one half, and sixty green tokens, the expected count is fifty and the standard deviation is five, giving z equal to two. This arithmetic is not by itself an authorship verdict: the threshold, assumptions, repeated contexts, and evaluation population matter. SynthID-Text uses a different published sampling design, with candidates competing through pseudorandom tournament scores. Google’s report documents its use for Gemini text, while its research and reference code describe scheme-specific detection. Longer, diverse generations can offer more evidence than short or tightly constrained answers, but length alone guarantees neither detection nor text quality. Editing or translation can weaken some watermarks. A negative detector result can also occur for machine-generated text produced without the tested watermark, so absence of that signal is not proof of human authorship.