Modelling Human Values for AI Reasoning

TL;DR

This paper tackles the challenge of embedding human values in AI by proposing a formal, computable Value Taxonomy Model (VTM) that grounds abstract values in verifiable properties and organizes them in multi-level taxonomies. It integrates social psychology theories (notably Rohan and Schwartz) to align the model with established concepts of values, value types, and value systems, while remaining agnostic to any single theory. A context-aware extension—context-based value taxonomies—allows values to evolve with time and situation, and a coherence/propagation mechanism maintains consistent value importance across the taxonomy. The approach is demonstrated through the uHelp running example and is validated against real-world domains (Hospital del Mar, firefighters, and community networks), illustrating practical pathways for value-aware decision making and value-aligned multiagent systems. The work also lays out a comprehensive roadmap for future research in value identification, aggregation, decision making, and MAS governance, aiming to enable provably value-aligned AI across diverse domains.

Abstract

One of today's most significant societal challenges is building AI systems whose behaviour, or the behaviour it enables within communities of interacting agents (human and artificial), aligns with human values. To address this challenge, we detail a formal model of human values for their explicit computational representation. To our knowledge, this has not been attempted as yet, which is surprising given the growing volume of research integrating values within AI. Taking as our starting point the wealth of research investigating the nature of human values from social psychology over the last few decades, we set out to provide such a formal model. We show how this model can provide the foundational apparatus for AI-based reasoning over values, and demonstrate its applicability in real-world use cases. We illustrate how our model captures the key ideas from social psychology research and propose a roadmap for future integrated, and interdisciplinary, research into human values in AI. The ability to automatically reason over values not only helps address the value alignment problem but also facilitates the design of AI systems that can support individuals and communities in making more informed, value-aligned decisions. More and more, individuals and organisations are motivated to understand their values more explicitly and explore whether their behaviours and attitudes properly reflect them. Our work on modelling human values will enable AI systems to be designed and deployed to meet this growing need.

Figures (8)

• Figure 1: Influencers of behaviour, and evaluating behaviour w.r.t. its alignment with values
• Figure 2: A selection of value definitions, adapted from ChengFleischmann2010Rohan2000. Text highlighted in yellow describes the nature of values, orange describes what values refer to or the issues they address, green describes the purpose of values or what they are for, blue highlights the notion of what is important or desirable, and pink describes who holds values or to whom they apply.
• Figure 3: Different levels of abstraction for the value fairness, with various grounding semantics: numbers indicate node importance
• Figure 4: Some value concepts may have more than one parent node
• Figure 5: uHelp's value taxonomy for the value fairness

Theorems & Definitions (5)

• Definition 1: Value taxonomy
• Definition 2: Coherence of value importance
• Proposition 1
• Proposition 2
• Definition 3: Context-based value taxonomy