Can AI Models Develop a Shared Sense of Ethics Through Networked Interaction?

Large language models (LLMs) are increasingly deployed in settings where multiple agents interact, raising questions about whether shared ethical behaviour can emerge collectively. Most alignment approaches focus on individual models, but human societies suggest that norms often develop through dialogue and persuasion.

This project investigates whether ethical convergence can arise among networked AI agents. Agents were modelled as nodes in a scale-free network, each assigned a moral vector across twelve principles. They responded to moral scenarios and reflected on neighbour reasoning, updating only if persuaded.

Analysis using cosine similarity, entropy, and principal component analysis showed that convergence is possible, though fragile. Polarisation often persisted, and biased agents, particularly those in central positions, strongly shaped outcomes. These findings complement recent research on emergent behaviours in LLM societies and highlight both the potential and risks of decentralised alignment.