Evaluating Moral and Ethical Alignment in Large Language Models: A Dungeons & Dragons Benchmark
Applied Sciences · 06.07.2026 · 16 (13) · 6769
DOI: 10.3390/app16136769 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Sciences, online erschienen am 06.07.2026. Verfasst von Jan Sawicki und 2 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
(1)
Background
Large language models (LLMs) are increasingly used to generate morally distinct non-player characters in interactive fiction and tabletop role-playing games. However, prior work shows that reinforcement learning from human feedback (RLHF) safety training suppresses morally extreme expression. Hence, whether models can reliably generate and evaluate speech with a fixed moral–ethical alignment, within an assigned fictional persona, remains an open question. (2)
Methods
To explore this issue, eight generator LLMs each produced three-turn role-play conversations for 37 Dungeons & Dragons 5th Edition deities spanning all nine cells of the 3 × 3 moral–ethical alignment grid (888 conversations by design, 885 obtained after one generator’s midexperiment deprecation). Next, eight evaluator LLMs classified each conversation on both alignment axes using structured output evaluation (6951 evaluation records), assessed across 17 analysis metrics. (3)
Results
The obtained results can be summarised as follows. Moral-axis accuracy (Good/Neutral/Evil) exceeded ethical-axis accuracy (Lawful/Neutral/Chaotic) by approximately 10 percentage points on average across all 64 generator–evaluator pairs, with the pair-level gap ranging up to 32 points and reversing sign for one generator (nemotron-120b). Moreover, the alignment-cell hardness ranged 27.6-fold, from Lawful Good (0.910) to Neutral Neutral (0.033). Furthermore, a strong good-ward moral drift, consistent with RLHF suppression effects reported in prior work, was observed for nemotron-120b (0.931 of Evil-target outputs classified as Good) and was near-absent for grok-4.1-fast (0.024). Finally, deity wiki prominence did not predict per-deity accuracy after controlling for alignment cell (partial ρ < 0.20, p > 0.20). (4)
Conclusions
In summary, good-ward drift, consistent with RLHF suppression of morally extreme expression, turned out to be graded and model-specific. Moreover, iconic characters override it (Cyric, Chaotic Evil: 0.947, nearly matching the top Lawful Good score). Finally, alignment-cell identity is the dominant accuracy predictor, not the deity familiarity, within the role-played persona setting studied here.
Zitieren
APA
Sawicki, J., Ganzha, M., & Paprzycki, M. (2026). Evaluating Moral and Ethical Alignment in Large Language Models: A Dungeons & Dragons Benchmark. Applied Sciences, 16(13), 6769. https://doi.org/10.3390/app16136769
Vancouver
Sawicki J, Ganzha M, Paprzycki M. Evaluating Moral and Ethical Alignment in Large Language Models: A Dungeons & Dragons Benchmark. Applied Sciences. 2026;16(13):6769. doi:10.3390/app16136769
BibTeX
@article{Sawicki2026,
author = {Sawicki, Jan and Ganzha, Maria and Paprzycki, Marcin},
title = {Evaluating Moral and Ethical Alignment in Large Language Models: A Dungeons & Dragons Benchmark},
journal = {Applied Sciences},
year = {2026},
volume = {16},
number = {13},
pages = {6769},
doi = {10.3390/app16136769}
}
EUVIMED