There is now a way to watermark text from diffusion language models, a newer class of AI text generators that had slipped past existing detection tools.
Most watermarking schemes for AI-generated text were built for autoregressive models, the standard type that writes one token after another in order. Those schemes lean on already-generated tokens to decide how to bias the next one. Diffusion language models break that assumption: they can generate tokens in any order, so the context a watermark needs is often incomplete or not yet decided. A paper posted to arXiv describes a fix: apply the watermark in expectation across the context even when some tokens are still undetermined, and favor tokens that strengthen the watermark for other tokens around them. The detector itself stays the same as in prior schemes. In testing, the method hit a true positive rate above 99% with little quality loss and held up about as well against tampering as existing autoregressive watermarks.
This matters because diffusion language models are moving from research curiosity to shipped product, and until now they had no working provenance tool. Autoregressive watermarking, from Google DeepMind's SynthID to various academic schemes, has become a reference point for AI-content labeling debates and policy proposals. A gap like this one is exactly the kind of blind spot that undercuts those efforts if it goes unaddressed.
Watermarks only work if every model family gets one and if providers actually turn them on - a technical fix here doesn't settle either of those open questions.