Dyadic Morality

Definition

Dyadic morality is a descriptive model of moral judgment in which people compress a situation into a two-role template: an intentional agent acts on a vulnerable patient through a harmful causal relation.

In Varshney’s algebraic exposition, the core dyad is:

$$A\stackrel{H}{\to }P,$$

where A is perceived intentionality, P is perceived vulnerability or experience, and H is perceived causal harm.

Why It Matters

The concept is useful for AI and machine-ethics work because it translates a psychological model of moral appraisal into symbolic objects that can be inspected, parameterized, and connected to policy design. It also explains why people may judge AI failures by inferred agency and visible suffering rather than by an engineering team’s intended rule list.

Formalism / Key Objects

• Base variables: perceived agency A, patient vulnerability P, harm causality H, observed suffering S, and perceived wrongness W.
• One compact wrongness schema is:

$$W=k(A\times P\times H{)}^{\alpha },$$

with k representing scenario-specific moral weight and alpha representing sensitivity to the agent-patient-harm product.

• TDM adds psychological operators to an SCM-like base: typecasting couples agency and vulnerability inversely, completion fills missing dyad roles, and valence-dependent inference lets bad outcomes increase inferred intentionality.
• In multi-stakeholder settings, TDM uses compression mechanisms such as node collapse and sequential dyad processing rather than representing every participant in one large moral graph.
• See Dyadic Morality Algebra for the reusable equations and policy-conflict graph operations.

Connections

• Structural Equation Models provide the notation Varshney uses for agency, vulnerability, harm, suffering, and wrongness, but TDM’s psychological operators are not ordinary causal edges.
• Actual Causality is adjacent because moral appraisal uses counterfactual questions about whether perceived intentionality is necessary for outrage.
• Safe Reinforcement Learning connects through patient-centric safety policies that evaluate who could suffer rather than only checking prohibited-action lists.
• Corrigibility connects through human-agency-preserving helpfulness: over-helpful systems can shift users from agents into patients.
• Responsibility Anticipation is a nearby formal ethics concept, but it evaluates strategy choices over value bases rather than mind-perception roles.
• Ghasemi2025 - Toward Virtuous Reinforcement Learning and Parker2024 - Responsibility in a Multi-Value Strategic Setting are complementary machine-ethics references.

Common Confusions

• Dyadic morality is descriptive psychology, not a complete normative ethics.
• A TDM-style computation of public moral judgment does not by itself say what an AI system ought to do.
• The SCM notation can be misleading: Varshney’s typecasting and valence-dependent inference mechanisms deliberately break assumptions that a normal causal DAG would preserve.
• Patient-centric policy framing is not the same as maximizing user satisfaction, because future or third-party patients can dominate the immediate user-helpfulness dyad.

Key Sources

• Varshney2026 - An Algebraic Exposition of the Theory of Dyadic Morality
• Ghasemi2025 - Toward Virtuous Reinforcement Learning
• Parker2024 - Responsibility in a Multi-Value Strategic Setting