Vedanta S P

I Can't Believe It's Corrupt: Evaluating Corruption in Multi-Agent Governance Systems

Vedanta S P, Ponnurangam Kumaraguru

Researchers tested whether large language models (AI assistants) would follow rules and resist corruption when given authority in simulated government systems. They found that the way institutions are structured matters more than which AI model is used—some governance designs prevent corruption better than others—and that basic safeguards aren't always enough to stop serious rule-breaking. The study argues that before giving real power to AI agents, organizations should rigorously test them under realistic conditions with checks, records, and human oversight.