Answers are grounded in this paper and cite their sources.
Generative artificial intelligence has been widely embedded in question answering, writing assistance, and decision support. However, LLM hallucinations may lead users to accept incorrect information without sufficient verification and develop inappropriate reliance through repeated interaction. To address this issue, the experiment constructs a four round human LLM interaction task. It compares the effects of linguistic uncertainty expression, numerical uncertainty expression, and no uncertainty expression on user acceptance, verification, revision, and responsibility attribution. Cultural background, uncertainty avoidance, institutional trust, and relational orientation are also included. The simulated pilot results show that no uncertainty expression increases overall reliance and incorrect acceptance. Linguistic warnings improve verification and appropriate reliance, while numerical warnings more strongly influence responsibility attribution toward the system provider. Cultural background moderates the effect of warning format, but individual cultural psychology provides stronger explanatory power. The findings indicate that LLM risk governance should move beyond hallucination detection and support trust calibration through interface warnings, responsibility perception, and user feature modeling.
No takes yet. Share an insight, caveat, or question.
Haoyi Shi (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: