Values as dynamic priors: neurobiological constraints on formal models of value alignment
Formal models of human values have brought much-needed rigor to AI alignment research by making value relations explicit and evaluable by machine. Yet the assumptions that make such models computationally tractable may limit their ability to capture how values are represented, regulated, and expressed in human cognition. Drawing on evidence from neuroscience and the predictive processing framework, we examine three assumptions common to many formal value models. First, value importance is typically treated as fixed within a given context, whereas neurobiological evidence suggests it is dynamically shaped by arousal, uncertainty, and learning. Second, hierarchical value taxonomies assume that values can be fully represented through grounding components, yet some higher-level values, including dignity, integrity, and authenticity, appear difficult to evaluate through weighted aggregation alone. Third, formal models typically restrict value influence to the evaluation of actions, whereas neurobiological evidence suggests that values also influence attention, perception, and the interpretation of situations before explicit evaluation begins. We argue that these observations do not undermine formal value modeling but rather identify its limits and opportunities for extension. Specifically, we propose representing uncertainty about value importance, and its change over time, through distributional estimates, distinguishing compositional from non-compositional values, extending value influence to the construction of situation representations, and incorporating bottom-up learning into value taxonomy development. More broadly, we argue that engagement with neuroscience can help distinguish productive simplifications from potentially limiting assumptions, providing a more empirically grounded basis for the development of value-aligned AI systems.
Authors
- Mark d’Inverno (ORCID: https://orcid.org/0000-0001-8826-5190)
- Nardine Osman (ORCID: https://orcid.org/0000-0002-2766-3475)
- Joydeep Bhattacharya (ORCID: https://orcid.org/0000-0003-3443-9049)
Institutions
- Hong Kong Baptist University (HK)
- Artificial Intelligence Research Institute (ES)
Publication Details
- Journal
- AI and Ethics
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1007/s43681-026-01528-y
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Hong Kong Baptist University