A new approach called 'reverse game theory' aims to control powerful AI models by designing systems that incentivize them to follow rules. As AI capabilities grow, ensuring alignment with human values is critical to prevent rogue behavior. Traditional methods like interpretability and constitutional AI have limitations, as models can hide intentions or underperform.
Reverse game theory, or mechanism design, focuses on creating rules that encourage desired behavior regardless of internal workings. Inspired by real-world examples like cake-cutting and auction systems, this approach uses cross-checking between AI agents. A recent paper by Stephen Morris and colleagues suggests rewarding models based on how well their answers predict another model's output, making deliberate underperformance a losing strategy.
Other methods include competitive dynamics and pairing biased agents. Elad Hazan from Princeton advocates for mechanisms inspired by human society, such as alternating roles between 'solvers' and 'auditors' to enforce accountability through mutual oversight.
Source: MIT Technology Review AI · Summarized by HeadlinesBriefing