Psychology and Social Cognition Conversational AI Systems Language Understanding and Pragmatics

Why do AI agents misalign with what users actually want?

UserBench explores how often AI models fully understand user intent across multi-turn interactions. The study reveals that human communication is underspecified, incremental, and indirect — traits that challenge current models to actively clarify goals.

Note · 2026-02-23 · sourced from Design Frameworks
Why do AI conversations reliably break down after multiple turns? How should researchers navigate LLM reasoning research?

UserBench evaluates agents in multi-turn, preference-driven interactions where simulated users start with underspecified goals and reveal preferences incrementally. The results quantify a gap that existing benchmarks obscure:

The framework identifies three core traits of human communication that make this hard:

  1. Underspecification — users initiate requests before fully formulating their goals
  2. Incrementality — intent emerges and evolves across interaction turns
  3. Indirectness — users obscure or soften their true intent due to social or strategic reasons

These are not edge cases — they are the default condition of human communication. Language is inherently ambiguous (Clark, 1996; Liu et al., 2023), and meaning is co-constructed through interaction.

The disconnect between task completion and user alignment is the critical finding. Standard benchmarks measure whether an agent completes a task — UserBench measures whether the agent completed the right task, from the user's perspective. Current models are task-capable but not user-aligned.

This connects to Why can't users articulate what they want from AI? — the 20% figure quantifies the double gap. And since How do users actually form intent when prompting AI systems?, the incrementality trait confirms that intent-as-binary is a design error, not an edge case.

The finding that models elicit <30% of preferences through active querying connects to Can models learn to ask clarifying questions instead of guessing? — proactive questioning is trainable (0.15% → 73.98%) but is not standard in current deployments.


Source: Design Frameworks

Related concepts in this collection

Concept map
16 direct connections · 117 in 2-hop network ·medium cluster

Click a node to walk · click center to open · click Open full network for a force-directed map

your link semantically near linked from elsewhere
Original note title

agents fully align with all user intents only 20 percent of the time — even best models elicit fewer than 30 percent of preferences through active querying