Lexical Dependence in Mental Health Detection Models
September 21, 2026
Researchers found that frontier models achieved an F1 of 0.846 for anxiety detection on Reddit but often relied on keyword matching. Testing on text with keywords removed revealed significant lexical dependence, suggesting models may not be capturing underlying linguistic patterns of anxiety.
HOW THIS AFFECTS YOU
●
builderBe cautious when deploying mental health classifiers that may fail if users stop using specific trigger words.
●
researcherYou must account for lexical shortcuts when evaluating fine-tuned classifiers on clinical or psychological tasks.