Authority, Truth, and Citation Bias: A Large-Scale Multi-Domain Benchmark for Studying Epistemic Susceptibility in Large Language Models

2026-06-11 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

AuthorityBench, a new 220,564-prompt multi-domain benchmark, reveals that citation presence significantly influences epistemic behavior in large language models (LLMs), often increasing hallucination rates. This benchmark employs a novel 2x2 factorial design, crossing claim veracity with citation veracity across general knowledge, science, law, and medicine domains, while controlling for 40 prompt templates, four venue prestige tiers, and author demographics. Evaluating seven models, the study found that both real and fabricated citations consistently elevate hallucination compared to a no-citation baseline. The effect is most pronounced when fabricated citations accompany true claims, boosting hallucination by 3 to 22 percentage points and reaching 35 to 77% in general knowledge, though legal claims demonstrate greater robustness. Venue prestige and author demographics had negligible impact.

Key takeaway

For machine learning engineers deploying LLMs in citation-augmented environments, you must implement robust fact-checking mechanisms, as this research demonstrates that citation presence, even fabricated ones, significantly increases hallucination rates. Do not assume citations guarantee factual accuracy; instead, prioritize independent verification, especially for general knowledge domains where susceptibility is highest. Your evaluation pipelines should account for this epistemic susceptibility.

Key insights

Large language models exhibit epistemic susceptibility, increasing hallucination rates when citations are present, regardless of their veracity.

Principles

• Citation presence consistently increases LLM hallucination.
• Fabricated citations with true claims amplify hallucination significantly.
• Legal domain claims show comparative robustness to citation bias.

Method

AuthorityBench uses a 2x2 factorial design, crossing claim and citation veracity across four domains, with controlled variation in prompt templates, venue prestige, and author names.

In practice

• Use AuthorityBench to evaluate LLM epistemic susceptibility.
• Prioritize fact-checking LLM outputs even with citations.

Topics

• Large Language Models
• Citation Bias
• LLM Hallucination
• Epistemic Susceptibility
• Benchmark Datasets
• Fact-Checking

Code references

• floating-reeds/AuthorityBench

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Visualizing and Benchmarking LLM Factual Hallucination Tendencies via Internal State Analysis and Clustering — Fabricated citations significantly amplify LLM hallucination, especially in more robust models like GPT-4o-mini.
• Trust, but Don't Verify: Epistemic Blind Spots in LLM Source Evaluation — LLMs can detect fabricated statistics but fail to apply this capability during multi-source synthesis, prioritizing stylistic credibility.
• Where Fake Citations Are Made: Tracing Field-Level Hallucination to Specific Neurons in LLMs — LLM citation hallucination is field-specific, with distinct neural mechanisms for different bibliographic components.
• Beyond Literal Summarization: Redefining Hallucination for Medical SOAP Note Evaluation — Lexical faithfulness in LLM evaluation misclassifies valid clinical reasoning as hallucination, inflating error rates.
• HalluCiteChecker: A Lightweight Toolkit for Hallucinated Citation Detection and Verification in the Era of AI Scientists — HalluCiteChecker detects and verifies AI-generated hallucinated citations to improve scientific integrity and reduce reviewer burden.
• Building Reliable Long-Form Generation via Hallucination Rejection Sampling — Rejecting hallucinated segments and resampling during generation mitigates snowballing in long-form LLM outputs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.