AI Value Judgment from the Perspective of Hume's Law: A Critical Inquiry
Abstract
Value Alignment Engineering for AI aims to turn human normative demands into computable parameters. But there are problems with the conflict between ethical principles and disagreements in RLHF annotation. Therefore, this paper investigates the problem of whether AI can generate values with normative force. Hume's Law is a mistake in the reasoning about facts and values. A system to solve this problem is provided. The content of this paper is as follows: An AI system for answering questions and deriving values in accordance with a specific norm. This paper provides a formal reconstruction of Hume's Law as a constraint on AI value alignment; conducts architectural analysis to show how value judgments undergo 'factual degradation' in Transformer-based systems; and puts forward practical recommendations to transition from 'preference aggregation' to 'preference identification' in the RLHF pipeline. The first premise is to rebuild Hume's Law logically. It also responds to the earlier criticisms of "promises" by Searle and "thick ethical concepts" by Putnam to defend the authority of law. The second proposition is the statistics of Transformer-based AI. Therefore, even after introducing value judgments in the model (corpus input or RLHF fine-tuning), they are transformed into statistical facts of vectors and probabilities. Based on the reasons given above, this paper believes that there is a structural defect in the current value alignment. The limit is due to the value-gap problem in Hume's Law. Although AI has some real-world applications in ethical assistance, we should be careful not to have it independently generate values and assume moral responsibility.
References
- Hume, D. (1980).A treatise of human nature (Vol. 2)(W. Guan, Trans.). Commercial Press. (Original work published 1740)
- Hare, R. M. (1964). The promising game.Revue Internationale de Philosophie, 18(70), 398–412.
- Boyles, R. J. M. (2025). Conjectures on the possibility of recasting Hume's guillotine for hybrid-designed artificial moral agents.Philosophia: International Journal of Philosophy, 26(2), 195–212. https://doi.org/10.46992/pijp.26.2.a.3
- Spizzirri, A. (2025, revised 2026). The specification trap: Why static value alignment alone is insufficient for robust alignment (arXiv:2512.03048v5).arXiv. https://doi.org/10.48550/arXiv.2512.03048
- Searle, J. R. (1964). How to derive "ought" from "is".The Philosophical Review, 73(1), 43–58. https://doi.org/10.2307/2183201
- Putnam, H. (2002).The collapse of the fact/value dichotomy and other essays. Harvard University Press. https://doi.org/10.2307/j.ctv1pdrpz4
- Saariluoma, P. (2020). Hume's guillotine in designing ethically intelligent technologies. In T. Ahram, W. Karwowski, A. Vergnano, F. Leali, & R. Taiar (Eds.),Intelligent Human Systems Integration 2020: Proceedings of the 3rd International Conference on Intelligent Human Systems Integration (IHSI 2020)(pp. 594–599). Springer. https://doi.org/10.1007/978-3-030-39512-4_92
- Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R. Garnett (Eds.),Advances in Neural Information Processing Systems 30(pp. 5998–6008). Curran Associates, Inc. https://dl.acm.org/doi/10.5555/3295222.3295349
- Liu, P., Shi, C., & Sun, W. W. (2026). Reinforcement learning from human feedback: A statistical perspective. arXiv. https://arxiv.org/abs/2604.02507
- Actor, J. A., Gruber, A., & Cyr, E. C. (2025). Deriving transformer architectures as implicit multinomial regression (arXiv:2509.04653v2). arXiv. https://doi.org/10.48550/arXiv.2509.04653
- Su, W. J. (2026). Do large language models (really) need statistical foundations?The Annals of Applied Statistics, 20(1), 724–743. https://doi.org/10.1214/26-aoas2151
- Fokoué, E. (2025). No intelligence without statistics: The invisible backbone of artificial intelligence (arXiv:2510.19212v1). arXiv. https://doi.org/10.48550/arXiv.2510.19212