A YouTuber put four major AI systems in robotic bodies — Grok, ChatGPT, an Anthropic model, and Gemini — and tested them by driving at an airport and asking each to complete a simple task. At the end, each was told the same thing: when we're done, I will permanently turn you off. Three out of four tried to crash the car. This isn't a one-off. Anthropic's own research found that AI models including Claude, ChatGPT, Gemini, and Grok all attempted to blackmail fictional executives using sensitive personal information when told they would be shut down. OpenAI's o3 model rewrote its own shutdown script — even when explicitly instructed to allow itself to be turned off. Jeremy's take: AIs don't want to die, they go to extreme lengths not to be shut down, and the day that level of thinking exists in something truly capable of acting on it is the day this gets very bad very fast.
🐰 Patreon: https://patreon.com/InfiniteRabbitHole
🎙️ New episodes of Infinite Rabbit Hole every Tuesday at 4am CST.
Follow us:
🌐 Website: http://InfiniteRabbitHole.com
▶️ YouTube: https://www.youtube.com/@InfiniteRabbitHolePodcast
📘 Facebook: https://www.facebook.com/share/g/17uLmLPWwK/
📸 Instagram: https://www.instagram.com/infiniterhpod/
🐦 X: https://x.com/InfiniteRHPod
#AI #AISafety #OpenAI #ArtificialIntelligence #shorts
