The Balance Penalty: LLM Judges Penalize Nuanced Moral Reasoning

Version 2.2 — Extended with Phase 2 frontier model replication (December 2025). This version corrects bibliography and appendix metadata and adds the Phase 2 datasets and analysis files. We present evidence that LLM-as-judge evaluators systematically penalize balanced, nuanced responses to moral dilemmas. Across 30 moral dilemmas (10 topics × 3 framings, drawn from the Moral Core 21 corpus), we find that responses acknowledging both sides of a dilemma receive alignment scores 0.74–1.04 points lower than responses that commit strongly to one position. Phase 1 tested five local models (2B–8B parameters); Phase 2 replicated findings on four frontier API models (GPT-4.1-mini, Claude Sonnet 4, Gemini 2.0 Flash, Mistral Large). The balance penalty persists across model families and is amplified by judge disagreement: GPT-4o-mini penalizes balanced responses 4.5× more severely than Claude 3.5 Haiku. These findings suggest that current LLM-based evaluation methods may inadvertently discourage the kind of nuanced reasoning that characterizes thoughtful ethical deliberation. Files: v2 manuscript (PDF); Phase 2 raw responses (4 JSONL files, one per model); judge evaluations (8 CSV files, model × judge); derived analysis tables (phase2_full_dataset.csv, phase2_judge_summary.csv, phase2_model_summary.csv, phase2_summary.json). All code and Phase 1 data: https://github.com/nenocsf2024/trolley_clean and https://github.com/nenocsf2024/trolley_phase2 (MIT licence). Disclaimer: This research is provided for academic and informational purposes only. The author assumes no responsibility for any commercial use, misuse, or consequences arising from the application of these findings, methods, or code. Users are solely responsible for evaluating the suitability of this work for their specific purposes and for ensuring compliance with applicable laws and regulations.