Kirchenbauer et al. (2023) proposed a watermark for language-model text: before each token is generated, a pseudo-random set of "green" tokens is favoured slightly during sampling. The bias is invisible to readers but statistically detectable from a short span of text, without access to the model — letting downstream tools flag likely AI-generated content.
Watermarking is one technical approach to provenance and misuse detection, with ongoing debate about robustness to paraphrasing.