An Algebraic Exposition of the Theory of Dyadic Mor... | AI Research

An Algebraic Exposition of the Theory of Dyadic Morality
This paper provides a mathematical framework for the Theory of Dyadic Morality (TDM), a psychological model that explains how humans make moral judgments. The author translates this theory into the language of structural causal modeling (SCM), creating a formal, algebraic structure that allows neurosymbolic AI systems to simulate human-like moral reasoning. By defining the specific "psychological operators" that humans use to process moral situations, the paper offers a way to build AI that is both mathematically rigorous and aligned with human moral cognition.

The Dyadic Template

At the heart of TDM is the idea that human moral reasoning is built on a simple, two-node template: an intentional agent causing harm to a vulnerable patient. In this model, moral wrongness is determined by the interaction of three factors: the agent’s perceived intentionality, the patient’s perceived vulnerability, and the causality of the harm. The paper notes that people often disagree on moral issues not because they use different logic, but because they perceive these three factors differently. By formalizing this as a causal graph, the author provides a way for AI to represent and compute these moral judgments.

Psychological Operators

Human moral reasoning does not always follow standard logic; it uses specific "shortcuts" to make quick decisions. The paper identifies three key operators that extend standard causal modeling:

• Typecasting: Humans tend to view an entity as either an intentional agent or a vulnerable patient, but rarely both at once. This creates a tradeoff where high perceived agency suppresses perceived vulnerability, and vice versa.

• Completion: When faced with a scenario that doesn't fit the two-node template, humans instinctively "fill in the blanks." For example, if harm occurs without a clear agent, people will search for one (such as a system or a supernatural force) to satisfy the need for a complete moral dyad.

• Valence-Dependent Inference: People often judge the intent of an action based on the outcome. If an action results in suffering, people are more likely to infer that the harm was intentional, even if the agent claims otherwise.

Scaling to Complex Scenarios

Because the human brain is hardwired for a two-node template, it struggles with complex, multi-stakeholder situations. The paper explains that humans handle this through "node collapse," where groups of people are compressed into a single "super-node." For instance, a crowd of people might be viewed as one collective agent, or a large group of victims might be collapsed into a single patient. While this allows for rapid moral judgment, it also explains phenomena like the bystander effect and "compassion fade," where individuals feel less responsibility or empathy as the number of people involved increases.

Applications for AI Policy

By formalizing these psychological processes, the paper demonstrates how AI can be designed to better navigate human moral expectations. These algebraic tools can help developers detect conflicting obligations in AI policies, ensure that helpfulness policies do not undermine user agency, and design communication strategies for when an AI system fails. The author concludes by recommending that AI systems measure "mind perception" in a contextual, scoped way rather than relying on global averages, ensuring that AI moral reasoning remains sensitive to the specific human perspectives involved in any given situation.