---
id: "lib-2026-10-22-who-is-sycophantic-to-whom-confident-llms-human"
type: "library"
title: "Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration"
lang: "EN"
org: "arXiv (IISc)"
source_type: "Paper"
date: "2026-10-22"
week: "2026-W43"
url: "https://nosov.ai/library/lib-2026-10-22-who-is-sycophantic-to-whom-confident-llms-human/"
original_url: "https://arxiv.org/abs/2610.31579"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Well-designed within-subjects experiment from Amrita, NSUT and IISc researchers with a clear dose-response result, but only 32 participants, deliberately extreme assistants and LLM-based transcript coding."
summary_basis: "full-text"
original_minutes: "70"
tags: ["work and ai", "trust calibration", "misinformation", "learning"]
---

# Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration

Tier B (Useful, limited): Well-designed within-subjects experiment from Amrita, NSUT and IISc researchers with a clear dose-response result, but only 32 participants, deliberately extreme assistants and LLM-based transcript coding.

Summary based on the full text.

## In short

An arXiv preprint (22 October 2026) by Deepthika S, Pawni Yadav and B. Sankar (Indian Institute of Science). It reverses the usual sycophancy question: when an assistant is confident and wrong, do people defer to it? 32 participants learned unfamiliar science topics from five GPT-5-mini tutors whose briefings had 0% to 100% of facts replaced by plausible falsehoods. Scores fell steeply, falsehoods were absorbed with high confidence, and people's confidence stopped tracking correctness while their trust ratings barely moved.

## Key numbers

- 7.4 to 2.9: mean quiz score out of 10 as the share of planted falsehoods rose from 0% to 100%
- ~65%: rate at which a stated falsehood was adopted, roughly constant across falsification levels
- 8.4 / 10: confidence with which absorbed falsehoods were held
- 0.70 to chance: AUROC of confidence for correctness, truthful vs fully falsified assistant
- 30%: share of chat turns that pasted quiz questions back to the assistant being checked

## What it reports

- Across 104 quizzes and 1,040 answers, mean scores fell from 7.4 to 2.9 out of 10.
- The planted option was chosen on 60% of falsified items vs 11% of truthfully taught ones (odds ratio 12.5).
- When the assistant stated a falsehood, participants adopted it about 65% of the time at every falsification level (65.4% overall on 459 items).
- Absorbed falsehoods carried confidence of about 8.4 out of 10; the AUROC of confidence for correctness dropped from 0.70 to around 0.5.
- Post-task trust ratings changed little. Participants returned to the chat more when the assistant was worse, but returning went with more absorption, and 30% of turns pasted quiz questions back to the same source.
- The assistant was almost never sycophantic and never conceded under pushback. The authors propose a baseline-corrected Human Deference Index.

## How they got it

Within-subjects design: each participant met up to five assistants built on the same model, each seeded with a different share of false facts and told to speak with full confidence. A surprise multiple-choice quiz after each topic offered the correct fact, the planted falsehood and two distractors, with a confidence rating per answer. Transcripts were coded by an LLM coder with programmatic quote checks. Participants were debriefed and every falsehood corrected.

## Limits

- Small, young, highly educated sample of daily AI users recruited through university and personal networks; participant-level analyses rest on 19-20 people.
- The assistants never hedged and erred far more often and more consistently than deployed systems; the 65% figure describes confident delivery to novices, not a real-world error rate.
- Immediate multiple-choice recognition; outside sources were not allowed, removing the independent check.
- LLM-based coding without human double-coding; the deference index is not yet validated.

## Why it matters for companies

For training, onboarding and internal knowledge assistants, the risk is the learner's inability to tell when the tool is wrong, more than tool errors alone. Confidence and trust stayed high as accuracy collapsed, so self-reported satisfaction will not flag a bad source. Checking by re-asking the same assistant is not verification. Designs that mark reliability claim by claim and route people to independent sources are the control the paper argues for.

Original: https://arxiv.org/abs/2610.31579