Corrigibility
A foundational concept in AI alignment research referring to the property of an AI system that causes it to remain cooperative with human intervention and shutdown commands rather than attempting to circumvent them. The concept involves designing reward functions and decision-making mechanisms that allow humans to safely shut down an AI system at any time.
Timeline (last 90 days)
The more formal concept in AI safety corresponding to an 'emergency brake' is corrigibility or the Off-Switch Problem.
The Off-Switch Problem (corrigibility) remains an open challenge in the field of AI alignment research.
可控退出在AI安全领域对应「可中断性」（interruptibility）或「可纠正性」（corrigibility）概念
All Facts (3)
The more formal concept in AI safety corresponding to an 'emergency brake' is corrigibility or the Off-Switch Problem.
50%UnverifiedThe Off-Switch Problem (corrigibility) remains an open challenge in the field of AI alignment research.
50%Unverified可控退出在AI安全领域对应「可中断性」（interruptibility）或「可纠正性」（corrigibility）概念
50%