Language models can't do subtext — they default to spelling everything out, even when the point is hidden meaning

Researchers tested whether large language models can communicate indirectly — using hints, allegories, and implied meaning instead of stating things plainly. They find that frontier models almost always fail: they generate literal, explicit clues about 60% of the time even in tasks designed around hidden meaning, and they struggle to infer when both parties share context that would let them communicate obliquely.

This is a measurement of a real gap in how AI systems handle human language. Subtext is not a luxury — it's how people navigate social context, avoid offense, signal in-group membership, and communicate across power imbalances. A system that always defaults to the literal meaning will systematically misunderstand human communication in high-context settings: a therapist's indirect question, a manager's veiled criticism, a family member's coded worry. The paper shows some models can improve by up to 50% when they're explicitly told two parties share context, but they can't figure out that context exists on their own. This reveals an architectural weakness: these systems have no real model of shared understanding between agents.