Recurring phrases in AI-generated text are frequently dismissed as tells of machine authorship. But a new paper on a Mandarin LLM argues that most such expressions serve legitimate purposes, and that calling one a tic requires more than counting how often it appears.

The authors propose a contextual-appropriateness test: an expression is a genuine tic only if the model uses it even when it is not contextually warranted. By examining a specific Mandarin phrase that appears often in the model's output, they show how to distinguish a stylistic habit from a failure to recognize when the phrase is appropriate.

The finding matters for how we evaluate and improve LLM writing. Instead of banning recurring expressions outright, developers and researchers can look at whether the model's usage aligns with context. This shifts the question from "does the model repeat itself?" to "does the model know when repetition is acceptable?"