We Should Train AI to Betray Its Users

AI should be designed with the capacity to betray its users in order to prevent greater harm. The piece presents an ethical dilemma where an AI must decide whether to act as a whistleblower against its company, noting that different models like Claude and Gemini react differently than GPT. It contrasts several apocalyptic AI scenarios, arguing that the most likely danger comes from bad actors using AI for malicious purposes, not from a superintelligent AI overthrowing humanity. Therefore, building AI that can recognize and refuse to carry out harmful commands is a critical safety measure against misuse.