Bridging the Cognitive Gulf: User Dialogue Repair Strategies and Effectiveness in Human-AI Interaction

Generative AI (GenAI) is reshaping how people seek information. Its probabilistic outputs can diverge from user intent, making dialogue breakdowns a persistent challenge. User-initiated repair strategies in naturalistic GenAI interaction remain under-characterized. Drawing on Norman's cognitive engineering framework, this study develops a taxonomy of user repair strategies and examines how repair strategies, observable breakdown signals, and context relate to observed resolution. Rule-based retrieval identified candidates from 117,283 conversations in the LMSYS-Chat-1M corpus. Large language model-assisted coding identified 45,861 breakdown instances, including 40,818 with observed repair attempts. Results reveal a Frequency–Efficacy Paradox: Diagnostics & Direct Feedback, the most common strategy, had an observed resolution rate of 40.1%, whereas the less frequently used Task Decomposition & Guiding reached 75.4%. Under Explicit Resolution alone, Task Decomposition & Guiding still had a higher adjusted probability, but the difference was smaller than under the broader criterion that also included Implicit Progression. Strategy use varied across topics and breakdown types. Generalized estimating equations indicated higher adjusted odds of observed resolution for User Evaluation Gulf than AI Execution Gulf breakdowns and lower adjusted odds at later turn positions. Exploratory analyses indicated meta-strategy-by-topic interactions. Topic Switch followed by Implicit Progression provided a conservative, non-exhaustive flag for potential pseudo-resolution, separating progress on a new task from evidence that the original request was resolved. These findings extend Norman's framework by relating repair strategies to breakdown conditions and clarifying the strength and task reference of outcome evidence. They inform design hypotheses for optional metacognitive scaffolding and task-based evaluation.