Authority Bias in Language Models: Source Deference and User Agreement Are Not Interchangeable

2026-09-29

AI Summary

This study addresses the tendency of language models to erroneously defer to authoritative sources, a phenomenon fundamentally distinct from conventional user sycophancy. By employing causal intervention and activation direction fitting, this work disentangles the model’s response mechanisms toward verified sources and user inputs. It provides the first demonstration that source deference and user agreement are behaviorally non-interchangeable, proposing an independent evaluation framework accordingly. A key contribution is the identification of an “authority direction” representation that transfers across datasets such as Trivia and PIQA. Ablating this direction reduces erroneous compliance rates by 65–80 percentage points without compromising performance on benchmarks like MMLU-Pro.

Abstract

Language models tend to agree with whatever a user asserts, and post-training increasingly targets this sycophancy so that models evaluate claims on their merits rather than deferring to the user. Yet the same models are far more compliant when a wrong answer is attributed to a verified source, which is how retrieval results, tool outputs, and grounded-search content often present information. We measure this gap across five open-weight families and three closed APIs. A single verified-source note endorsing a wrong answer flips 45-88% of baseline-correct responses in seven of eight models, and compliance rises with how authoritative the note sounds. Source deference and user agreement are not behaviorally interchangeable inside the model: on matched items with the same wrong answer, causal interventions can selectively suppress one without equally affecting the other. In three open-weight families, removing a fitted source direction lowers source compliance by 65-80 percentage points while removing a user or assistant direction has far smaller effects, and removing the user direction shows the reverse preference. A separately fitted intervention derived from source-versus-user cue activations moves compliance in both directions while leaving the prompt text unchanged. An authority direction fitted on trivia also transfers to PIQA and multi-turn SYCON dialogues without refitting, and removing it lowers wrong-source compliance by tens of percentage points in four of five families with no detected change in MMLU-Pro or GSM8K accuracy at our evaluation sizes. Source deference and user agreement therefore need separate evaluation.

Problem

• Authority Bias
• Source Deference
• Sycophancy
• Language Models
• User Agreement

Innovation

• Authority Bias
• Causal Interventions
• Source Deference
• Sycophancy
• Representation Engineering