Do language models overestimate how often irony appears?

GPT-4o can interpret ironic intent in emoji usage. But it systematically overestimates ironic intent compared to humans — the median irony score assigned by GPT-4o is significantly higher than human perception (p < .001). LLMs detect irony as a category but miscalibrate its prevalence (Irony in Emojis: A Comparative Study of Human and LLM Interpretation).

This overestimation reveals something important about how LLMs process pragmatic meaning. Irony detection is a pattern-matching success: the model has learned which textual features correlate with ironic intent in its training data. But ironic patterns are over-represented in training data relative to their actual frequency in human communication, because ironic usage is more salient, more commented upon, more explicitly labeled than sincere usage. The model learns the pattern but not the base rate.

This is a specific instance of a broader calibration problem. Since Why do preference models favor surface features over substance?, we know that training data artifacts systematically distort model judgments across multiple dimensions. Irony overestimation is the pragmatic version: the model's sense of "how often is this ironic?" is calibrated to training data saliency, not to real-world frequency.

The implication for literary analysis is significant. Literary irony is subtle, context-dependent, and often operates through understatement — exactly the opposite of the salient, explicitly marked irony that dominates training data. A model that over-reads ironic intent will find irony where an author intended none, and may miss genuine irony that operates through restraint rather than exaggeration. Since Can language models adapt implicature to conversational context?, the failure to calibrate irony to context is part of a larger pattern: LLMs apply fixed pragmatic templates where communicative context should modulate interpretation.