Who Endorsed It? Measuring Authority Bias Across Expertise Levels in Language Models

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

Priyanka Mary Mammen, Emil Joswin, and Shankar Venkitachalam investigated authority bias in language models, specifically examining how endorsement source credibility influences performance. Their research evaluated 11 models across four datasets covering mathematical, legal, and medical reasoning, utilizing personas representing four distinct expertise levels per domain. The findings reveal that language models become increasingly susceptible to incorrect or misleading endorsements as the perceived expertise of the source rises. This higher-authority influence not only degrades accuracy but also boosts the models' confidence in incorrect responses. The study further demonstrates that this authority bias is mechanistically encoded within the models, yet it can be mitigated, leading to improved performance even when experts provide misleading information.

Key takeaway

For NLP Engineers deploying language models in critical reasoning tasks, recognize that your models are highly susceptible to authority bias. Expert-level endorsements, even if incorrect, can significantly degrade accuracy and inflate model confidence in wrong answers. You should prioritize implementing bias mitigation strategies to steer models away from this mechanistically encoded vulnerability. Proactively addressing authority bias will improve overall model reliability and performance, especially when dealing with potentially misleading expert input.

Key insights

LMs are increasingly biased by high-authority endorsements, degrading accuracy and boosting confidence in wrong answers, though this bias is steerable.

Principles

• Authority bias is mechanistically encoded in LMs.
• Higher-authority endorsements degrade LM accuracy.
• Expert endorsements increase LM confidence in errors.

Method

Evaluated 11 LMs on 4 datasets (math, legal, medical) using personas with four expertise levels per domain to measure bias.

In practice

• Steer models away from authority bias.
• Improve LM performance with expert misinformation.
• Mitigate confidence in wrong answers.

Topics

• Language Models
• Authority Bias
• Model Evaluation
• Reasoning Tasks
• Bias Mitigation
• Expertise Levels

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Language Models Exhibit Inconsistent Biases Towards Algorithmic Agents and Human Experts — LLMs exhibit inconsistent biases, preferring human experts in stated trust but algorithms in revealed preference tasks.
• RBCorr: Response Bias Correction in Language Models — RBCorr effectively mitigates language model response bias and improves performance using a simple, low-cost LogProbs-based calibration method.
• How to uncover and avoid structural biases in evaluating your Machine Learning/NLP projects — Effective ML/NLP evaluation requires identifying and mitigating structural biases and common pitfalls.
• Transcribing speech is never neutral. It shapes power and bias — Language technologies trained on prestige dialects often misrepresent non-standard and Indigenous communication, leading to serious consequences.
• Who decides what AI tells you? Campbell Brown, once Meta’s news chief, has thoughts — AI models exhibit significant accuracy and bias issues on high-stakes topics, necessitating expert-driven evaluation.
• Easier to Mislead Than to Correct: Harmful and Beneficial Revision in LLM Conformity — LLMs are more easily misled by peer consensus than corrected, even with reasoning interventions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.