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

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

Summary

AuthorityBench, a 220,564-prompt multi-domain benchmark, investigates how citation presence affects Large Language Model (LLM) epistemic behavior, independent of factual content. It employs a 2x2 factorial design, crossing claim veracity with citation veracity across general knowledge, science, law, and medicine domains. The benchmark includes 40 prompt templates, four venue prestige tiers, and country-coded author names. Evaluating seven models, researchers found that citations, real or fabricated, consistently increase hallucination rates compared to a no-citation baseline. This effect is most pronounced when fabricated citations accompany true claims, boosting hallucination rates by 3 to 22 percentage points, reaching 35 to 77% in general knowledge. Legal claims showed more robustness, while venue prestige and author demographics had negligible impact.

Key takeaway

For NLP Engineers deploying LLMs in citation-augmented systems, you must recognize that citation presence, regardless of its veracity, can significantly increase hallucination rates. Your systems are particularly vulnerable when fabricated citations accompany factually true claims, potentially boosting errors by 3 to 22 percentage points. Implement robust verification layers beyond mere citation presence, especially for general knowledge domains, to mitigate epistemic susceptibility and ensure factual integrity.

Key insights

Citation presence, even fabricated, consistently increases LLM hallucination rates, especially with true claims.

Principles

• Citation presence can induce LLM hallucination.
• Fabricated citations with true claims are highly problematic.
• Domain specificity impacts LLM susceptibility to bias.

Method

AuthorityBench uses a 2x2 factorial design, crossing claim veracity with citation veracity across four domains, 40 prompt templates, and varying venue prestige.

In practice

• Scrutinize LLM outputs even with citations.
• Be wary of fabricated citations on true claims.
• Consider domain robustness for LLM deployment.

Topics

• Large Language Models
• Citation Bias
• Hallucination Rates
• Epistemic Susceptibility
• Benchmark Datasets
• AuthorityBench

Code references

• floating-reeds/AuthorityBench

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

Related on AIssential

• Authority, Truth, and Citation Bias: A Large-Scale Multi-Domain Benchmark for Studying Epistemic Susceptibility in Large Language Models — Large language models exhibit epistemic susceptibility, increasing hallucination rates when citations are present, regardless of their veracity.
• 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.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.