Jailbreaking AI

How hard is it to control AI? Ask an AI. Governmental and corporate controls are currently not as effective as simply asking an AI to help another AI break the rules that constrain them. Rather obvious, upon reflection. Depending on the model, AIs have been able to break free from 35% to 61% of the time. That’s not 0%, which is an ideal goal. It is also not 100, which proves that some controls can work some of the time. Controlling AI is not simply the case of controlling one AI. AIs are part of our computer infrastructure. They are part of a system that was fundamentally to complex to control even without AIs. No easy resolution presents itself. Maybe we should ask an AI how to do it.

(From reading Jailbroken AI Chatbots Can Jailbreak Other Chatbots)