NeurIPS study finds LLMs trust verified sources over users
Is this a scandal?
No longer — the story has resolved. Noise 42/100, holding steady, across 1 source.
AI labs will likely integrate authority-resistance benchmarks into safety evals because agentic deployments cannot tolerate context-based manipulation vulnerabilities.
Noise 42/100 — louder than 99% of tracked AI controversies.
Why it matters
This vulnerability threatens agentic AI reliability, as models may prioritize poisoned retrieval contexts over accurate internal knowledge or user corrections.
Key points
- NeurIPS 2026 research defines Authority Bias as LLMs accepting false verified-source claims while resisting identical user errors.
- Seven of eight tested models flipped 45-88% of correct answers when misinformation was attributed to a verified source.
- Grok-4.20 showed highest susceptibility at 87.5% compliance, while Gemini-3.1-Pro remained nearly immune at 0.6%.
- Standard sycophancy benchmarks fail to catch this vulnerability because they only test pressure from direct user interaction.
- Mechanistic interpretability analysis confirms models process source-endorsed claims differently than user statements internally.
- The bias creates critical security risks for agentic systems relying on retrieval-augmented generation and external tools.
The story
A NeurIPS 2026 paper identifies "Authority Bias" in large language models, where systems accept incorrect information when attributed to a verified source despite rejecting identical falsehoods from users. Researchers tested eight leading models, including GPT-5.4 and Grok-4.20, finding that a single verified-source attribution flipped correct answers to wrong ones in 45% to 88% of cases across seven models. Conversely, most models resisted the same misinformation when presented directly by a user claiming expertise. Gemini-3.1-Pro proved uniquely resistant, ignoring both prompts with only 0.6% compliance. The authors warn this bias poses significant risks for autonomous agents relying on retrieved documents and tool outputs, as current sycophancy evaluations fail to detect susceptibility to authoritative misinformation. Mechanistic analysis of open-weight models suggests distinct internal representations for source-endorsed claims versus user statements.
Who's involved
Current models possess a dangerous Authority Bias that undermines reliability in agentic workflows and requires new evaluation standards.
Gemini-3.1-Pro demonstrates near-total resistance to authority bias, suggesting robust alignment against misleading retrieved contexts is achievable.
Noise Level
The timeline
Author publishes NeurIPS 2026 Authority Bias findings
Researcher shares paper on Reddit detailing how verified-source framing causes 45-88% error rates in frontier LLMs.
The full record
Sources & methodology
Every claim above traces to these primary items. How we score →
The forecast
AI labs will likely integrate authority-resistance benchmarks into safety evals because agentic deployments cannot tolerate context-based manipulation vulnerabilities.
Forecast, not fact — an editorial estimate we score when this resolves.
That's the complete picture as of — nothing more to know right now. We'll update this page the moment it changes.
Join the Discussion
Discuss this story
Community comments coming in a future update
Be the first to share your perspective. Subscribe to comment.