What large language models know and what people think they know
Authors: M Steyvers, H Tejeda, A Kumar, C Belem
Published: 2025
Publication: Nature Machine ..., 2025 - nature.com
LLMs often lead to user overestimation of response accuracy, especially with longer explanations; adjusting explanation styles to align with model confidence improves calibration and discrimination gaps, enhancing trust in AI-assisted decision making.
Methods: Conducted experiments using multiple-choice and short-answer questions to study user confidence versus model-stated confidence; varied explanation length and alignment with model internal confidence.
Key Findings: Calibration gap (human vs. model confidence), discrimination gap (ability to distinguish correct vs. incorrect answers), and effects of explanation style and length on user trust.
Limitations: Specific details about participant sample size and demographic diversity are not provided; broader applicability across varied user populations remains unexamined.
Institution: University of California Irvine
Research Area: Computational Linguistics, Computational Social Science,AI Ethics, Trust in AI
Discipline: Computational Social Science
Citations: 100