The Loophole framework turns a user's stated moral principles into a detailed rule system, then sends two adversarial agentsA multi-agent system contains multiple AI agents that interact, coordinate, divide work, or influence one another while pursuing tasks. looking for failures. One searches for immoral actions that remain legal under the generated rules, while the other finds moral actions that the rules prohibit. A judging agent patches straightforward translation errorsAgent self-correction is an AI agent's ability to detect a problem in its work and revise its actions without restarting the entire task. and escalates genuine ambiguity or contradiction to the userHuman-in-the-loop AI keeps a person involved in reviewing, correcting, approving, or guiding an AI system's work..
The demonstration begins with DNA privacy and uses synthetic cases to expose missing boundaries in a moral code. The same pattern could test chatbot constitutions by searching for both forbidden answers that slip through and appropriate answers the system refusesCounterexample generation searches for a specific case that disproves a rule's claimed coverage or exposes an unintended outcome., or compare personal data preferences with a company's terms and surface disagreements before a contract is accepted.
More speculative experiments model legislators and synthetic voter personas, then test or revise proposed bills against their inferred preferences. The presenter frames these government applications as aspirational and explicitly notes privacy and logistical problems, so the work is best understood as an exploratory stress-testing methodAdversarial evaluation deliberately searches for inputs and edge cases that reveal where an AI system's rules or behavior fail. rather than a validated representation of public values.
Watch on YouTube
