EthicaAI

At a Glance

Abstract

EthicaAI formalizes Amartya Sen's "Meta-Ranking" theory — preferences over preferences — as a dynamic mechanism in Multi-Agent Reinforcement Learning. We demonstrate that Situational Commitment — morality conditional on survival — is the only Evolutionarily Stable Strategy across 4 environments, 7 SVO conditions, and up to 1,000 agents.

Key Findings (v2)

• Dynamic meta-ranking (λ_t) significantly enhances collective welfare (p=0.0003, Cohen's f²=0.40)
• Agents exhibit emergent role specialization — Cleaners vs Eaters (? divergence p<0.0001)
• Only "Situational Commitment" survives as ESS (~12% of population, 200-gen replicator dynamics)
• Individualist SVO (θ=15°) best matches human PGG data (WD=0.053)
• SVO rotation accounts for 86% of total effect (full factorial 2³ decomposition)
• Human-AI behavioral alignment: Wasserstein Distance < 0.2 across all SVO conditions
• Byzantine robustness: cooperation maintained with up to 50% adversarial agents
• Scale invariance verified from 20 to 1,000 agents (ATE stable ±0.03)

Seven Stages of Research

• Stage 1: Core PGG with 7 SVO conditions and causal inference
• Stage 2: Mixed-SVO populations, communication channels, continuous action spaces
• Stage 3: Cross-environment validation (IPD, PGG, Harvest), evolutionary dynamics
• Stage 4: Human behavioral alignment, 100-agent scalability
• Stage 5: MAPPO training, partial observability, LLM comparison
• Stage 6: Vaccine allocation, AI governance, human-AI interaction
• Stage 7: Byzantine robustness, Moran process, GNN architecture, policy implications

Environments

• Cleanup — shared resource management dilemma
• Iterated Prisoner's Dilemma (IPD) — cooperation vs. defection
• Public Goods Game (PGG) — free-rider problem at scale
• Harvest — common-pool resource depletion

Tech Stack