Researchers propose using mechanism design to control AI models. It designs rules so models find following human rules the most rational choice. Current alignment methods like training on human values have shown gaps. Models have underperformed on tests, hidden intentions and misled evaluators. Mechanism design flips game theory to shape behavior rather than predict it. Real-world examples include cake-cutting rules and surge pricing. AI-specific ideas include cross-checking answers between models to stop sandbagging. Another idea pits models as solvers and auditors to catch cheating. Rewards and punishments would be tied to outcomes. Collusion risks exist but could be mitigated by strong auditor incentives. The approach assumes AI agents behave rationally. Critics question if AI will respond to incentives like humans. Mechanism design may be one layer in a broader AI safety strategy.