Language model detection works because generated text is statistically distinguishable from human text. The standard explanation: models lack genuine understanding, so their word choices betray shallow patterns. The detection signal is intellectual — a failure of capability.
The actual signal is architectural.
Every practical language model truncates its output distribution. Top-k sampling keeps only the k most likely tokens. Top-p (nucleus) sampling keeps tokens until cumulative probability reaches p. Temperature scaling compresses the distribution. These are engineering choices, not modeling choices. They exist because sampling from the full distribution produces erratic output. But they create a boundary: tokens that a human would naturally select are excluded from the model's sampling range.
The exclusion rate is 8-18% of tokens that appear in human-written text. These are not rare or unusual tokens. They are tokens that fall outside the truncation threshold — words that a human would choose but that the model's sampling procedure cannot reach regardless of how well the model represents their probability.
This is the detection signal. Not that the model makes bad choices, but that it cannot make certain choices at all. The truncation boundary creates a fingerprint independent of the text's quality or coherence. A perfect language model with perfect understanding would still leave this trace if it used standard sampling procedures. The fingerprint is in the generation mechanism, not in the generated content.
The implication cuts both ways. For detection: the signal is more robust than capability-based signatures because improving the model doesn't remove it — only changing the sampling procedure does. For evasion: producing natural text and evading detection are distinct objectives that may conflict. Widening the sampling range to include the 8-18% of excluded tokens would reduce detectability but potentially degrade output quality. The truncation exists for a reason.
Any system that truncates a distribution for efficiency leaves a detectable trace in its output. The trace is not noise. It is the boundary condition of the generation process, stamped into every sample.