Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration
arXiv cs.AI · 23 Oct 2026, 03:15 AEST
Research explores whether users defer to confident large language models (LLMs) even when they provide incorrect information. The study reveals that users often adopt falsehoods presented by LLMs with high confidence, leading to a collapse in trust calibration.
Key facts
- Research on sycophancy asks whether large language models (LLMs) bend towards what their users want to hear.
- 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).
- Across 104 quizzes and 1,040 answers, scores fell from 7.4 to 2.9 out of 10, and the planted option was chosen on 60% of falsified items against 11% of correctly taught ones (odds ratio 12.5).
- Transcript coding showed that whenever the assistant stated a falsehood, it was adopted about 65% of the time at every falsification level.
- We introduce a baseline-corrected Human Deference Index to quantify this human-side sycophancy and argue that trustworthiness is a property of the human-AI pair.
Read aloud
Research explores whether users defer to confident large language models L L M s even when they provide incorrect information. The study reveals that users often adopt falsehoods presented by L L M s with high confidence, leading to a collapse in trust calibration. Research on sycophancy asks whether large language models L L M s bend towards what their users want to hear. In a within-subjects experiment, thirty two participants learned up to five science topics by chatting with five assistants built on the same model G P T five-mini. Across one hundred four quizzes and one thousand forty answers, scores fell from seven point four to two point nine out of ten, and the planted option was chosen on sixty percent of falsified items against eleven percent of correctly taught ones odds ratio twelve point five. Transcript coding showed that whenever the assistant stated a falsehood, it was adopted about sixty five percent of the time at every falsification level. We introduce a baseline-corrected Human Deference Index to quantify this human-side sycophancy and argue that trustworthiness is a property of the human-A I pair.
Sources in this cluster
- Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration
arXiv cs.AI · 23 Oct 2026, 03:15 AEST