Reverse game theory tackles rogue AI
new approach uses incentives to control ai models, even if they dont understand them. this could improve ai safety and alignment.
a new approach called reverse game theory aims to control powerful ai models by creating incentives for them to follow rules. this method, also known as mechanism design, is borrowed from economics and tries to design rules that elicit desired behavior. researchers like stephen morris and elad hazan suggest mechanisms such as rewarding models for predicting each others outputs, pitting agents against each other, or pairing biased agents to reach optimal answers. this approach could improve ai safety and alignment, even if ai models internal workings are not fully understood.