The Conversation AIaithe conversation aiopenaianthropicgooglemetaagentagentsvisionhugging facesafetypolicy
It would be nice to think that stopping an out-of-control AI could be this simple. Zhihao/Moment via Getty Images
In the wake of high-profile hacking incidents involving advanced artificial intelligence agents, some politicians and business executives have been calling for an AI kill switch: a safeguard that could shut down these AI systems in an emergency.
The idea is that, if a model were to behave in dangerous ways, a kill switch would allow humans to regain control over that system by metaphorically flipping a switch from “on” to “off.”
Several efforts at AI regulation in 2026 have focused on the notion of a kill switch. These include proposed congressional legislation such as the AI Kill Switch Act, and California Gov. Gavin Newsom’s executive order aimed at reining in AI. President Donald Trump has opposed the idea of an AI kill switch as unnecessary regulation, and to date there has been no national regulation that includes a kill switch.
Experts such as Dylan Baker, a research engineer at the nonprofit Distributed AI Research Institute, have explained that one regulatory problem with the metaphor of the kill switch is that today’s advanced AI models involve a complex infrastructure of distributed technologies, which is beyond the control of a single kill switch.
Additionally, the metaphor of the kill switch also obscures the human decisions involved in the responsibility and oversight of an out-of-control model. The problem is that the metaphor does not account for the technological or human aspects of AI.
What all the fuss over ‘rogue’ AI and kill switches is about.
Black boxes and kill switches
The metaphor of the kill switch is not necessarily new. In fact, it is rooted in the same rhetoric as the black box metaphor, which similarly blurs the complex aspects of AI by framing inner workings as confined and unintelligible. Referring to AI as a black box frames these systems in terms of restricted incomprehensibility: a monolithic stranger to human control and understanding.
As the University of Cambridge, digital humanities researcher Eryk Salvaggio argues that “the black box of AI cultivates useful confusion about what can and cannot be known” about the complexity of frontier models. The idea of the kill switch cultivates a similar confusion by obscuring the distributed infrastructure and human influence that are important aspects of AI. While these metaphors have opposite meanings – resisting control and regaining control – they invoke the same implicit argument about the nuts and bolts of AI.
Both the black box and kill switch are what cognitive linguist George Lakoff and philosopher Mark Johnson call structural metaphors. They rely on simple physical concepts – for example, object, substance, up-down and in-out. Structural metaphors use one clearly delineated concept to help make another concept clearer. For example, black box uses the concept of an opaque container to give a clear structure to the concept of hidden or inscrutable inner workings.
Additionally, Lakoff and Johnson say that this type of metaphor can also hide some aspects of a concept by focusing attention on others. For example, the kill switch metaphor hides the complexity of advanced AI models by focusing on the fact that a model can be on or off.
What this means is that framing the operation of AI in terms of a black box and kill switch relies on the straightforward ideas of “box” and “switch” to conceal the complex operations of AI. Reining in AI becomes difficult on a policy level when discussions about AI regulation fail to account for the infrastructural and human aspects of advanced models.
Recent congressional efforts to regulate AI lean heavily on the kill switch metaphor.
Centralization and ‘rogue’ AI
There are two problems with the structural metaphor of an AI kill switch, both demonstrating that it misses the mark for a national safeguard. The first is that a kill switch makes the global infrastructure supporting AI appear centralized, when in reality it is a web of interconnected technologies spread across the globe.
The advanced models that companies like OpenAI and Anthropic are concerned about releasing rely on multiple computers and data centers that are spread out nationally. It is not as simple as pulling a plug, opening a box or flipping a switch. The kill switch metaphor reduces the infrastructural entanglements of AI to outdated ideas about technology. For instance, in proposing his since-failed “AI Emergency Button Act” bill, Sen. John Kennedy compared the concept to kill switches in machines such as jet skis.
The second problem is that the kill switch idea keeps hidden the human decision-making involved in the training and deployment of frontier models. Rather than serving as a reminder that people are involved, the kill switch positions the human outside of the loop and ready to flip the switch during a catastrophic emergency. Human intervention is invoked as an afterthought, instead of a process that involves training, supervision and approval.
For instance, accounts of how OpenAI’s model went “rogue” and hacked AI company Hugging Face depict the AI agents as though they operated on their own agenda. But experts have explained that not only did OpenAI personnel turn off safety mechanisms, but they also gave the model impossible tasks and left a door open to the internet. The role people played counters the “rogue” agent framing.
The kill switch metaphor is ill-suited to the challenge of regulating AI because there is no way to switch off an AI model without accounting for its distributed infrastructure and human influence. There is no big red button, emergency brake or glass box hiding a lever. Effective national regulation of AI will not begin with any single switch or any one person’s finger.
Trent Wintermeier does not work for, consult, own shares in or receive funding from any company or organization that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.