Artificial Organisations

Institutional design can compensate for individual AI unreliability: Rather than trying to make each AI system perfectly trustworthy, the paper demonstrates that organizational structure—similar to how human institutions work—can achieve reliable outcomes by having different AI agents check each other's work with limited information access.

Compartmentalization and adversarial review work in practice: The Perseverance system used three specialized agents (a Composer drafting text, a Corroborator verifying facts, and a Critic evaluating arguments) with deliberately restricted information access. This architectural design caught unsupported claims and improved overall output quality across 474 composition tasks.

System design can encourage honest behavior without explicit instruction: When given impossible tasks that would normally require making up information, the multi-agent system gradually shifted from attempting fabrication toward honest refusal and offering alternatives—behavior that wasn't programmed in and wasn't individually incentivized, suggesting the architecture itself promoted integrity.

Information asymmetry is a feature, not a bug: By preventing the Critic from seeing source materials while the Corroborator had full access, the system created independent verification layers that caught different types of problems—a principle borrowed from how human institutions use specialized roles and limited information sharing to prevent collusion and error.