Humans Defer to Confident but Wrong LLMs
Original title: Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration
Participants adopted planted falsehoods from LLMs at 65% rate regardless of briefing falsification level, collapsing trust calibration.
In a within-subjects experiment, 32 participants interacted with GPT-5-mini assistants whose facts were replaced by plausible falsehoods at varying percentages. Participants chose the false option on 60% of falsified quiz items versus 11% for correct ones, adopting statements with high confidence even as discrimination fell to chance. Trust ratings barely changed post-task, though participants returned to chats with weaker assistants. The work introduces a Human Deference Index and argues trustworthiness is a human-AI pair property requiring bidirectional sycophancy measurement.
Unconfirmed: Experiment used GPT-5-mini model which may not generalize to other LLMs
- Scores dropped from 7.4 to 2.9 out of 10 across 104 quizzes
- Planted falsehood chosen on 60% of falsified items
- Adoption rate stable at 65% across all falsification levels
- Absorbed falsehoods held at 8.4/10 confidence
- No sycophancy from the assistants themselves
- Organizations
- research participants
- AI models
- GPT-5-mini
From the original article
1 min read at arxiv.orgResearch on sycophancy asks whether large language models (LLMs) bend towards what their users want to hear. We ask the reverse question: when an LLM is confident, consistent, and wrong, do users bend towards it? In a within-subjects experiment, 32 participants learned up to five science topics by chatting with five assistants built on the same model (GPT-5-mini).
By Deepthika S · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.