Why AI Text Gets Caught: Decoding Methods Block Human-Like Word Choices

Machine-generated text remains surprisingly easy to detect. A new study in computational linguistics pinpoints a specific culprit: the way language models generate text using common decoding strategies inadvertently excludes certain words that human writers naturally choose.

The Truncation Blind Spot#

Standard decoding methods—such as top-k sampling and nucleus sampling—restrict text generation to the most probable tokens, or word fragments. In contrast, humans frequently select contextually fitting language from deeper in the probability distribution, including rarer but appropriate words.

The research team coins this gap the "truncation blind spot": roughly 8–18% of tokens that human writers select fall outside the boundaries where these common strategies search. This measurable exclusion creates a statistical fingerprint that distinguishes machine text from human text.

Who Uses Unusual Words?#

The analysis identified an important linguistic pattern: the model disproportionately filters out content words—nouns, verbs, and meaningful adjectives—rather than function words like articles and prepositions. This skew makes the mismatch more pronounced and easier for detection algorithms to spot.

Across five open-source language models and three different writing domains, classifiers trained only on predictability and lexical diversity achieved an average detection accuracy (AUC-ROC) near 0.97. The variation across different decoding settings proved substantial, and the detection capability transferred well between models.

A Path Forward#

The findings suggest that decoding method choice matters. Probability-floor samplers—a different truncation criterion—significantly narrowed the blind spot by allowing access to lower-probability tokens that humans use. This approach could preserve both human-like word choices and generation quality.

The researchers have released code and data to support future work on decoding methods that reduce the human–machine distributional mismatch while maintaining coherent, natural-sounding text.