No single ‘kill switch’ can shut down AI, experts warn

The notion of a "kill switch" to shut down artificial intelligence has resurfaced as concerns over the technology's potential dangers continue to grow. However, AI experts warn that no single solution exists to completely turn off AI systems.

Prof. Nicolas Papernot of the University of Toronto explains that a kill switch in AI refers to the ability for the person or organization that deployed the AI to halt its actions by denying it access to the necessary computing power. But this becomes challenging if the program has operated beyond its initial environment.

AI expert Thierry Poibeau from the CNRS argues that the image of a simple button to "switch off AI" is misleading, as there are numerous companies worldwide, each with their own software and services. There's no single entity that controls AI as a whole.

However, it is possible to shut down a specific AI system if one controls the machines and services it depends on. AI systems can be classified into three complexity levels, with the first being most controllable. In scenarios where AI is decentralized and given permissions to interact with various systems, shutting everything down becomes much more difficult.

The experts also caution that the implementation of a kill switch could have significant repercussions. As AI becomes more integrated into daily life, shutting it down could disrupt essential services in businesses, public administrations, and healthcare. The costs and drawbacks of such a move would be enormous.

Some high-profile figures, like US President Trump and Meta CEO Zuckerberg, have dismissed fears about AI's existential risks, arguing that market incentives and legal liability are sufficient safeguards. However, AI professor Jean-Gabriel Ganascia argues that the real risk lies in our increasing dependence on AI systems, rather than the possibility of an AI developing a mind of its own.