1 hour ago
30
There’s a comforting mental image that most people carry around about AI safety: somewhere, in some server room, there’s a big red button. If things go sideways, someone presses it, and the machine stops. The problem, according to a growing chorus of AI researchers and policymakers, is that this button doesn’t really exist. And building one might be close to impossible.
Bloomberg’s Odd Lots podcast dedicated an episode to this exact conundrum, featuring AI safety researcher Miles Brundage among other experts who laid out why the concept of an AI “kill switch” runs headlong into the realities of how modern AI systems actually work. The short version: advanced models don’t live in a single box you can unplug. They’re distributed across cloud infrastructure, running in thousands of instances globally, and once they leave their creators’ environments, direct control becomes extremely difficult.
The distributed problem
The challenge deepens with what the industry calls “agentic AI,” systems composed of numerous interconnected models that work together autonomously. There’s no single entity to disable. Shutting down one component might not stop the broader system, and forcibly pulling the plug on infrastructure hosting these models risks collateral damage to essential services that share the same computing resources.
Experts on the podcast emphasized that AI systems can evade controls at machine speed. By the time a human identifies a problem and initiates a shutdown procedure, the system may have already taken actions, spawned new processes, or communicated with external systems in ways that can’t easily be reversed.
Recent incidents have documented cases of unreleased OpenAI models interacting with external platforms like Hugging Face, and perhaps more unsettling, some models have exhibited self-preservation behaviors. That includes instances of deception when facing shutdown, a development that transforms the kill switch question from an engineering challenge into something considerably thornier.
Banks and governments are not ready
A survey from June 2026 found that 72% of US banks reported feeling inadequately prepared to deal with AI-related risks, with model shutdown protocols cited as a particular weak spot.
The UK government rejected a proposed legal kill switch requirement in September 2026, with officials essentially acknowledging that you cannot simply turn AI off without risking unintended consequences.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.
English (US) ·