Source-linked AI summary
The simulation of judgment in LLMs
Edoardo Loru, Jacopo Nudo, Niccolò Di Marco, Alessandro Santirocchi, Roberto Atzeni, Matteo Cinelli, Vincenzo Cestari, Clelia Rossi-Arnaud, Walter Quattrociocchi
TL;DR
The paper examines how LLMs operationalize evaluative judgment and whether apparent agreement with human and expert ratings reflects similar reasoning. It benchmarks six models against expert systems and human participants using a shared evaluation protocol, finding aligned outputs alongside different criteria and pattern-based mechanisms. These differences are associated with political asymmetries and a risk that plausibility can replace verification in delegated judgment.
Problem
As LLMs enter evaluative processes, it remains unclear how their judgments are produced and how their heuristics differ from human evaluation.
Method
The paper compares six LLMs with NewsGuard, MBFC, and human participants using news domains and a shared procedure for selecting criteria, retrieving content, and producing justifications.
Results
LLM outputs often align with expert ratings, but models and humans prioritize different reliability criteria, with models relying on lexical associations, statistical priors, and structural cues.
Takeaways & Limitations
Delegating evaluation to LLMs can shift how reliability and bias are operationalized toward statistical approximation, creating risks of political asymmetry and plausibility replacing verification.
Takeaways & Limitations
The sample size limits generalizability, despite the controlled design and multilingual robustness.
Abstract
from arXiv · showhide
Large Language Models (LLMs) are increasingly embedded in evaluative processes, from information filtering to assessing and addressing knowledge gaps through explanation and credibility judgments. This raises the need to examine how such evaluations are built, what assumptions they rely on, and how their strategies diverge from those of humans. We benchmark six LLMs against expert ratings--NewsGuard and Media Bias/Fact Check--and against human judgments collected through a controlled experiment. We use news domains purely as a controlled benchmark for evaluative tasks, focusing on the underlying mechanisms rather than on news classification per se. To enable direct comparison, we implement a structured agentic framework in which both models and nonexpert participants follow the same evaluation procedure: selecting criteria, retrieving content, and producing justifications. Despite output alignment, our findings show consistent differences in the observable criteria guiding model evaluations, suggesting that lexical associations and statistical priors could influence evaluations in ways that differ from contextual reasoning. This reliance is associated with systematic effects: political asymmetries and a tendency to confuse linguistic form with epistemic reliability--a dynamic we term epistemia, the illusion of knowledge that emerges when surface plausibility replaces verification. Indeed, delegating judgment to such systems may affect the heuristics underlying evaluative processes, suggesting a shift from normative reasoning toward pattern-based approximation and raising open questions about the role of LLMs in evaluative processes.
Introduction
As LLMs become embedded in evaluative workflows, the paper asks how delegated judgment is operationalized and how model procedures differ from human reasoning. It benchmarks model evaluations against expert and human judgments to examine the criteria and assumptions behind apparent output alignment.
- LLMs are increasingly used for classification, evaluation, recommendation, and decision support, embedding evaluative functions into sociotechnical systems.
- The central question is which heuristics LLMs encode when decisions are delegated, since similar outputs may conceal different underlying processes.
- The benchmark compares six LLMs, two expert rating systems, and human evaluators on reliability and political bias, using news domains as a controlled testbed.
- Credibility and bias judgments shape information exposure, public trust, belief formation, and public debate, making their operationalization societally consequential.
- Models often align with expert ratings, but structured comparisons show different reliability criteria and a shift from context-dependent normative reasoning toward pattern-based approximation.
Results and Discussion
Across benchmark and controlled evaluations, LLMs often align with expert reliability ratings but exhibit political asymmetries and recurring linguistic markers. Their classifications remain broadly accurate for clear-cut cases while diverging from humans and raising questions about reliance on statistical associations rather than content-specific evaluation.
- Evaluation framework: The benchmark evaluates six LLMs against expert ratings by comparing classifications, political labels, explanatory keywords, and structured evaluation procedures.The models are assessed using standardized prompts and, in the agentic workflow, procedural steps shared with human evaluators.
- Benchmark results: For MBFC ratings, models often exceed 90% accuracy at Low and High credibility extremes but diverge on Medium credibility sources.GPT-4o mini and Llama 4 Maverick classify 75% and 77% of Medium-credibility domains as unreliable, while Gemini produces a more balanced distribution.
- Political asymmetries: Classification errors are politically asymmetric: reliable Right-leaning outlets are more often labeled unreliable, while Center and Left-leaning outlets tend to have reliability overestimated.The passage attributes this pattern cautiously to correlations in training data rather than establishing partisan preference.
- Linguistic heuristics: Models use recurring linguistic markers, linking reliable domains with neutrality, transparency, factual reporting, and local news, and unreliable domains with misinformation, conspiracy, bias, and political terms.Keyword distributions are heavy-tailed, with a small core set of markers repeatedly associated with reliability judgments.
- Statistical associations: Even with only a domain URL, Gemini reaches an F1-score of 0.78 versus 0.86 with full HTML, while GPT reaches 0.77 versus 0.79.This pattern is consistent with assessments being shaped by prior knowledge about outlets rather than content-specific evaluation alone.
Conclusions
The study finds that LLMs can align with expert reliability and bias judgments while using evaluative mechanisms that differ from human reasoning. These differences include lexical and statistical pattern reliance, political asymmetries, and epistemic opacity in automated judgment.
- LLM outputs often align with expert classifications, especially when identifying unreliable sources, but this agreement can conceal divergent evaluative mechanisms.
- LLMs rely on lexical associations, statistical priors, and structural cues, with right-leaning outlets disproportionately classified as unreliable.
- Both humans and LLMs prioritize Factual Accuracy, but LLMs emphasize Ownership Transparency while humans give greater weight to stylistic and rhetorical features.
- The comparison suggests that LLMs partially replicate psychological behavioral regularities while operationalizing evaluation through fundamentally different mechanisms.
- Automated evaluative pipelines introduce epistemic opacity, making it important to examine how they operationalize normative categories such as reliability and bias.
- The controlled design provides a foundation for further research, but the study’s sample size limits the generalizability of its findings.
- Delegating evaluation to LLMs risks embedding lexical and statistical associations into judgment, potentially amplifying existing information pathologies.
Materials and Methods
The study evaluates news domains using standardized homepage inputs, zero-shot prompting, and an agentic workflow that separates criterion selection, content retrieval, and reliability assessment. A parallel human experiment follows the same structured procedure.
- The dataset comprises NewsGuard-rated English-language outlets selected from the US, UK, Canada, Australia, and New Zealand.
- Homepage HTML was filtered to remove scripts, styles, and navigation elements, retaining primarily textual content relevant to reliability assessment.
- All six LLMs used zero-shot, closed-book prompts without example classifications or definitions of reliability.
- The common prompt asks models to classify a homepage as reliable or unreliable using only its contents and to return the result in JSON.
- Model prompting: Queries for four models were truncated to the first 50,000 characters, affecting less than 2% of domains.
- Model prompting: Domains were evaluated individually to avoid bias from assessing reliability relative to a specific subset of domains.
- Agentic workflow: The workflow used sequential orchestration, randomized criterion presentation, content retrieval, independent criterion agents, and a final agent assigning an overall binary rating.
- Human experiment: A parallel human study recruited 50 participants for criteria selection and evaluation tasks in a controlled laboratory setting.