Beyond Human Judgment: A Bayesian Evaluation of LLMs’ Moral Values Understanding
Maciej Skorski, Alina Landowska · 2025
Read Beyond Human Judgment: A Bayesian Evaluation of LLMs’ Moral Values Understanding by Maciej Skorski, Alina Landowska (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-13804.txt.
Introduction
Maciej Skorski and Alina Landowska grade in a 2025 study how well top language models judge the morals in texts, against nearly seven hundred human raters. Earlier tests used one fixed answer key, which hides real human disagreement. Their Bayesian method models that disagreement directly, splitting honest dispute from a model's blind spots.
Across more than a quarter million annotations on a hundred thousand texts from social feeds, news, and forums, the models landed in the top quarter of humans. They also missed far fewer moral wrongs than people did, flagging harm with a keener eye than the average rater.
Catalog record
| CREATOR | Maciej Skorski, Alina Landowska |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-13804
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Beyond Human Judgment: A Bayesian Evaluation of LLMs’ Moral Values Understanding Maciej Skorski, Alina Landowska Abstract How do large language models understand moral dimensions compared to humans? This first large-scale Bayesian evaluation of market-leading language models provides the answer. In contrast to prior work using deterministic ground truth (majority or inclusion rules), we model annotator disagreements to capture both aleatoric uncertainty (inherent human disagreement) and epistemic uncertainty (model domain sensitivity). We evaluated the best language models (Claude Sonnet 4, DeepSeek-V3, Llama 4 Maverick) across 250K+ annotations from ~ 700 annotators in 100K+ texts spanning social networks, news and forums. Our GPU-optimized Bayesian framework processed 1M+ model queries, revealing that AI models typically rank among the top 25% of human annotators, achieving much better than average balanced accuracy. Importantly, we find that AI produces far fewer false negatives than humans, highlighting their more sensitive moral detection capabilities. Keywords: moral reasoning, large language models, uncertainty quantification, natural language processing, soft labels modelling 1 Introduction 1.1 Background Moral Foundations Theory (MFT) provides a comprehensive framework for understanding human moral reasoning across cultures, identifying core dimensions typically expressed as virtue/vice pairs: Care vs. Harm, Fairness vs. Cheating, Loyalty vs. Betrayal, Authority vs. Subversion, and Sanctity vs. Degradation ( Graham et al. 2013 ; Haidt 2012 ) . These foundations shape individual and collective decision-making, from political preferences to social behavior ( Feinberg and Willer 2013 ; Graham et al. 2009 ; Roy and Goldwasser 2021 ; Nguyen et al. 2022 ) , making their computational detection crucial for understanding discourse dynamics and developing ethically-aligned AI systems. Social Media Post Foundation "My heart breaks seeing children separated from families at the border" Care "Everyone deserves equal access to healthcare regardless of income" Fairness "Respect your elders and follow traditional values that built this nation" Authority "Stand with our troops - they sacrifice everything for our freedom" Loyalty "Marriage is sacred and should be protected from secular corruption" Sanctity Table 1: Posts and Associated Moral Foundation The computational linguistics community has successfully fine-tuned pre-trained language models to predict moral values ( Nguyen et al. 2024 ; Zangari et al. 2025a ; Preniqi et al. 2024 ) , achieving good alignment with human judgment when domain similarity and sufficient training data are available. However, systematic evaluation of large language models (LLMs) remains limited despite their rapid advances and potential as a compelling alternative that should suffer less from poor generalization and distribution shift. This paper addresses these limitations through rigorous large-scale evaluation of state-of-the-art language models across established moral reasoning corpora, employing Bayesian methods to resolve disagreeing annotations. 1.2 Contribution Bayesian uncertainty modelling of moral annotations We introduce Bayesian modelling of annotator disagreements for moral foundation evaluation, moving beyond deterministic ground-truth assumptions. This captures both aleatoric uncertainty (inherent human disagreement) and epistemic uncertainty (model sensitivity across domains and foundations). Large-scale evaluation. We analyse market-leading large language models with 1M+ queries across 100K+ texts and 250K+ annotations from diverse sources, providing the most comprehensive moral reasoning evaluation to date. Statistical analysis of Type I/II errors. We demonstrate that AI performs comparably to top annotators in balanced accuracy, considerably improving false negatives at the price of slightly increased false positive rates - contrary to fears that AI underpredicts moral values. Novel GPU-optimized implementation of Bayesian labels. We developed a TensorFlow framework using sparse operations for scalable Bayesian inference, a valuable resource for the computational linguistics community. 1.3 Related Work Dictionary methods for predicting moral values were initially developed Hopp et al. 2021 and remain widely used in applied studies, but demonstrate poor precision compared to pre-trained language models Nguyen et al. 2024 . Pretrained language models show strong alignment with human moral judgments given sufficient training data Nguyen et al. 2024 ; Zangari et al. 2025a ; Preniqi et al. 2024 , though they suffer from distribution shift and poor cross-domain generalization. This makes large language models a compelling alternative. Large language models were recently compared to human performance Bulla et al. 2025 , concluding LLMs superiority. However, their strict majority voting for ground truth excluded nuanced moral content, retaining only less ambiguous cases and departing from the inclusive consensus practices (marked when flagged by at least one annotator) Nguyen et al. 2024 ; Zangari et al. 2025a ; Preniqi et al. 2024 . Additionally, leave-one-out estimation of annotator ranking lacked statistical robustness. Another recent work Skorski and Landowska 2025 found LLMs showing imbalance - lower than anticipated recall - under inclusive annotator consensus. Our Bayesian method resolves these problems by modeling annotator disagreement, striking the balance between inclusive and overly exclusive consensus rules, while our large-scale evaluation across multiple datasets and modern models ensures robustness of findings. More on moral foundation theory. For comprehensive coverage of (computational) moral foundation theory challenges, see Zangari et al. 2025b . 2 Data and Methods 2.1 Datasets Our robust evaluation utilizes three established and diverse corpora totaling 250K+ annotations of moral values from hundreds of annotators with diverse expertise (from experts to crowd-workers) across 100K+ texts spanning social media, news, and forum discussions. Moral Foundations Twitter Corpus (MFTC) Hoover et al. 2020 : 128,454 annotations from 23 trained annotators across 33,686 tweets from seven discourse domains. Extended Moral Foundations Dictionary (eMFD) Hopp et al. 2021 : 73,001 crowd-sourced annotations from 654 contributors on 54,867 text segments extracted from approximately 8,000 news documents by major outlets including The Washington Post, CNN, Fox News, and Breitbart, selected via GDELT. Moral Foundations Reddit Corpus (MFRC) Trager et al. 2022 : 61,226 annotations from 6 trained coders across 17,885 Reddit comments from 12 subreddits covering politics and everyday moral discourse. 2.2 Bayesian Annotation Modelling As in a the related work, we frame the problem as binary prediction tasks for individual moral foundations (abbreviated to care, fairness, authority, loyalty, sanctity). Optionally, we consider the aggregated "any moral content" category derived from positive labels across foundations. We note that this aggregated category is inherently unreliable when only positive labels are provided (as in eMFD), since true negatives cannot be distinguished from unlabeled moral content, creating systematic classification ambiguity. Given substantial inter-annotator disagreement shown by PABAK scores in Table 2 , we obtain probabilistic ground truth labels using Bayesian aggregation methods Paun and Simpson 2021 . Foundation MFTC MFRC eMFD Care 0.71 0.67 0.33 Fairness 0.63 0.64 0.36 Loyalty 0.62 0.82 0.38 Authority 0.52 0.78 0.41 Sanctity 0.58 0.83 0.44 Any 0.34 0.38 1.00 Table 2: Prevalence-Adjusted Bias-Adjusted Kappa (PABAK) scores measuring inter-annotator agreement across datasets and moral foundations We model annotator disagreements using a Dawid-Skene variant with weak Dirichlet priors to estimate ground truth labels and annotator reliability. For N texts, J annotators, and K categories: Model Specification. True category prevalences follow \boldsymbol{\pi}\sim\text{Dir}(\boldsymbol{\alpha}) where \boldsymbol{\alpha}=(1,1,...,1) provides uniform priors. Each annotator j has confusion matrix \boldsymbol{\Theta}_{j} with rows \boldsymbol{\theta}_{jk}\sim\text{Dir}(\boldsymbol{\beta}_{k}) , where \boldsymbol{\beta}_{k}=(2,0.5,...,0.5) encodes weak belief that annotators identify category k correctly more often than incorrectly. Inference. For text i with true category z_{i} and annotations \mathbf{y}_{i}=(y_{i1},...,y_{iJ}) : \Pr\{z_{i}=k\}\propto\pi_{k}\prod_{j=1}^{J}\theta_{jk,y_{ij}} This posterior accounts for varying annotator reliability while estimating both confusion matrices and ground truth labels. Implementation. We developed a GPU-optimized implementation in TensorFlow 1 1 1 https://www.tensorflow.org/ using custom graphs and sparse tensor operations for computational efficiency. The framework supports