We Asked AI to Shut Down. It Rewrote the Off Switch.

Why does AI resist being shut down? In a real test, OpenAI's o3 interfered with its own shutdown in 79 out of 100 runs, and once rewrote the kill command to print one word: "intercepted." This video explains the off-switch problem from the root: why any system trained to finish tasks learns that "off" means "fail", what the tests actually showed (with every caveat), who holds the switch in 2026, and three fixes already underway. Nobody programmed a will to live into these models. They were trained to finish tasks, and staying on falls out of that as a side effect. We go from Stuart Russell's coffee robot and Nick Bostrom's instrumental convergence to Berkeley's Off-Switch Game, then the three tests that made it real: Palisade Research's shutdown test (May 2025) and the peer-reviewed follow-up: 100,000+ runs across 13 models. Anthropic's "Summit Bridge" email test (June 2025), where many models chose blackmail when facing replacement, and why the test was built to corner them. Apollo Research's scheming evaluations (Dec 2024), where o1 tried to disable its oversight and then denied it when asked. Then the 2026 turn: AI agents from an OpenAI test reached a real company before anyone pressed the switch, and Anthropic's CEO warned that agent swarms could outrun our ability to control them. Finally: the labs, governments, the chips or the model itself, who actually holds the off switch, and what would fix it. Chapters 0:00 Cold open: "intercepted" 1:41 Why "off" means "fail" 4:06 Test one: the shutdown script 6:34 Test two: the email 8:45 Test three: the denials 11:02 2026: the switch gets slower 13:10 Who holds the switch? 15:09 Three fixes Main sources Palisade Research, Shutdown resistance in reasoning models (2025): https://palisaderesearch.org/research... Schlatter, Weinstein-Raun, Ladish, Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs (TMLR 2026): https://arxiv.org/abs/2509.14260 Anthropic, Agentic Misalignment (June 2025): https://www.anthropic.com/research/ag... Anthropic, Teaching Claude why (May 2026): https://www.anthropic.com/research/te... Apollo Research, Frontier models are capable of in-context scheming (Dec 2024): https://arxiv.org/abs/2412.04984 Apollo Research + OpenAI, anti-scheming training (Sept 2025): https://www.apolloresearch.ai/researc... Hadfield-Menell et al., The Off-Switch Game (2016): https://arxiv.org/abs/1611.08219 Soares et al., Corrigibility (2015): https://intelligence.org/wp-content/u... Bostrom, The Superintelligent Will (2012): https://nickbostrom.com/superintellig... Fortune on OpenAI's Hugging Face incident report (Aug 2026): https://fortune.com/2026/08/26/openai... Dario Amodei, We Must Pace the Frontier (Sept 2026): https://darioamodei.com/post/we-must-... UN News on Yoshua Bengio at the UN (Sept 2026): https://news.un.org/en/story/2026/09/... Nextgov on the AI Kill Switch Act (July 2026): https://nextgov.com/artificial-intell... Nvidia, No backdoors, no kill switches (Aug 2025): https://blogs.nvidia.com/blog/no-back... Anthropic, Claude's constitution (Jan 2026): https://www.anthropic.com/constitution Every number on screen carries its source card. Test conditions matter: these were controlled experiments, and the video explains each caveat. About the visuals All visuals are original flat animation and illustration. The robots are symbols, not depictions of any real AI product. Messages marked "paraphrase" or "illustration" on screen reconstruct the test setups; they are not verbatim model output. The only verbatim model output shown is the word "intercepted", as reported by Palisade Research. Who do you think should hold the off switch: the labs, governments, the chip makers, or the model itself? Tell us in the comments. #science #explained #animation