What is an AI Kill Switch?

As the name suggests, an AI Kill Switch “kills” an LLM in case of a catastrophe. By baking these switches into their code, big tech companies have promised that even if all AI guardrails and weights fail, there will always be a way to shut down an AI from compromising human ethics and national security. 

Kill switches first originated at the Seoul AI Safety Summit, where tech giants such as Microsoft, Amazon, and OpenAI pledged that safety frameworks including AI kill switches would be published in order to prevent misuse of the technology (CNBC 2025). The summit created clear lines to define the risk associated with AI systems and how they would respond to threats such as bioweapons and cyberattacks. If the system could not guarantee the mitigation of these risks, the companies confirmed that they would use a kill switch in order to cease the development of their AI models.

Issues and Alternatives to Kill Switches

Recently, NVIDIA has started to push back on the notion that kill switches are optimal to mitigate any type of risk. David Reber Jr., NVIDIA’s Chief Security Officer, has argued that hard-coded, single-point controls would be a gift for hackers and would undermine the very global digital infrastructure that kill-switches aimed to protect (NVIDIA 2025). Reber has highlighted the paradox with kill switches, as they act as a safeguard but their very existence could introduce vulnerabilities and trust in US technology. In a digital AI race between the United States and China, these vulnerabilities could make or break who comes out on top. NVIDIA is also at the forefront of foreign AI chip development, meaning any policy misstep could inadvertently strengthen China and cause a problem for US national security.

Kill switches also present other problems including interstate confusion. What if an AI system that needs to be shut down in California but the servers are located in Singapore? Understanding whose laws apply in a given circumstance may prove to be more difficult and initially imagined. Kill switch laws can easily be bypassed by establishing servers in jurisdictions without heavy limitations in order to avoid shutdowns. These loopholes will prove to become an even greater threat in the context of the digital divide between the US and China, as developers may strategically relocate infrastructure in order to give geopolitical an edge.

The central problem to kill switches is the lack of nuanced decisions. Operating a kill switch boils down to just pressing a button or not. In order to allow for a variety of responses to imminent danger, we should instead move toward real-time monitoring and usage restrictions in order to create a spectrum of possible interventions. 

The Future of AI Governance Regarding Kill Switches

What the question of regulation on kill switches is starting to boil down to is who gets to decide when to implement them. Does NVIDIA have the authority to determine that kill switches are a national security issue? Or should that be left to countries and global treaties, who may not have industrial knowledge but do understand the political state of the world? 

Another point to consider is to look at other similar situations that have happened in history. When governments first grappled with the existential risks of nuclear weapons, which many feared could bring out the end of the world, they turned to international treaties and respective oversight mechanisms in order to balance control over the weapons. When the stock market crashed in 1926, 1974, and 1987, our government decided to use people to regulate tiered responses to the crashes rather than leaving stock autonomy solely to traders (Investopedia 2024). By taking inspiration from these previous disasters and the solutions that followed them, we can maybe get a better grasp as to how to handle AI kill switches.

As tech giants start to follow in the footsteps of NVIDIA and Jensen Hunag, AI ethicists alike should start to consider the potential benefits and vulnerabilities regarding kill switches and their role in keeping AI models and companies alike in check.