How to control the action options available to AI: Joe Carlsmith's alignment problem series, part seven

In the seventh article of the "How to Solve the Alignment Problem" series, Joe Carlsmith divided control over AI's optional actions into "limiting rogue options" and "other incentive shaping." The core challenge of limiting rogue options is that superintelligence may adopt "superhuman strategies" that humans cannot anticipate in advance or detect in real time, compounded by four sub-challenges: scarce supervision, delayed supervision, adversarial dynamics, and opacity. The response approaches are divided into controlling AI capabilities and controlling its environment (including supervision and environmental hardening), but effectively supervising superintelligent agents may itself depend on untrustworthy AI labor.