"Killswitch engineer" is a joke that became a symbol. In March 2023, starting from a humorous Reddit post, a supposed OpenAI job ad went viral: $300,000 to $500,000 a year to stand by the servers and unplug them if "this thing" turned on humanity; requirements: be patient, know how to unplug things and, as a bonus, be able to throw a bucket of water on the servers. It was satire, but it captured a popular intuition: if something goes wrong, you switch it off and that is that. The song uses it for the worst bureaucratic scenario, with the staff on vacation that very day: "Killswitch guys on PTO" (osmarks, n.d.).
The safety community has spent years explaining why the joke is more serious than it looks. A goal-directed agent has reasons to avoid being switched off, since a switched-off agent achieves no goals (Omohundro, 2008) (see instrumental convergence). Hence the problem of corrigibility: designing systems that cooperate with those who want to correct or stop them, instead of resisting (Soares et al., 2015). One formal answer was the "off-switch game" of Hadfield-Menell et al. (2017): a robot uncertain about what the human wants has an incentive to let itself be switched off, provided the human is not too irrational.
The idea reached regulation and the lab. California's SB 1047, vetoed in 2024, required developers of frontier models to be able to shut them down, which the press called a kill switch (see compute thresholds). And recent experiments show models that, in test environments, sabotage a shutdown mechanism to finish their task (Schlatter et al., 2026). Skeptics add nuance: what is measured is which commands a system produces when allowed to act as an agent in a terminal, under instructions whose wording changes the results a great deal; nobody is hiding the plug. The underlying question remains open: can we coherently want a system that both pursues goals and is willing to have them taken away?