Recommended
Is it possible to stop an artificial intelligence? Technically, an AI can be interrupted as long as it is in a controlled environment. But the more it acts autonomously across various systems, the more complex the shutdown becomes, experts warn.
The question seems increasingly relevant considering that AI-related security incidents are multiplying and fears about the dangers of this technology are on the rise.
- Is there an "emergency kill switch"? -
"In the field of AI-related cybersecurity," an emergency shutdown or "kill switch" refers to "the ability to interrupt the execution of an AI agent's action"—a program built on a generative AI model capable of chaining a series of tasks on its own—explains Nicolas Papernot, a professor at the University of Toronto and a specialist in computer security and AI, to AFP.
This mechanism would have to be able to be activated if an AI escapes all control and poses a safety risk.
An AI remains a "computer program," the researcher points out.
"The person or organization that created it can stop it by denying it access to the computing power necessary for its execution," he adds.
The situation becomes complicated when the program is authorized to act beyond its initial environment.
According to Papernot, this can happen if a "computer worm" (malicious program) controlled by AI spreads using the resources of other devices to duplicate itself. "In that case, all copies generated by the worm would have to be interrupted simultaneously, device by device," he explains.
For Jean-Gabriel Ganascia, a computer science professor at Sorbonne University, the question about the "kill switch" may feed the idea of "a metaphysical risk" with an AI that would have a will of its own. But for him, the more concrete danger is our growing dependence on these systems.
- Can an AI really be turned off? -
For Thierry Poibeau, director of research at the CNRS and an AI specialist, the image of a single button that allows one to "turn off the AI" is misleading.
"You cannot really cut off AI just like that, there are a huge number of companies around the world, of programs, of different services," he states. "There is no one who controls AI" as a whole.
On the other hand, it is technically possible to stop a specific system when you control the machines and services it depends on.
In an article published on LinkedIn, Hussein Abbass, a computer science professor at the University of New South Wales (UNSW) in Canberra, presents three levels of complexity.
In the first, the system remains in the environment controlled by its operator: implementing an emergency stop device is "in theory and in practice achievable."
In the second level, the agent has permissions to intervene in other programs, files, or machines: it becomes more difficult to stop it, especially actions it has already triggered.
In the third, the system extends its action to multiple systems, the AI becomes a free agent, and the goal is no longer so much to "turn it off" as to identify the elements that must be stopped before others.
"Once it has been set in motion, it is a bit like a computer virus," Poibeau concludes.
- What are the dangers of a "kill switch"? -
The more AI is integrated into daily activities, the more likely it is that its interruption will also block legitimate uses of this technology.
As we "automate more and more aspects of daily life thanks to AI," "it [will become] increasingly difficult to distinguish a bad AI from a harmless one and, therefore, to know what needs to be 'turned off'," explains Nicolas Papernot.
Stopping an AI system could result in the deterioration of services on which companies, governments, or health centers already depend.
"In a hypothesis of a real loss of control" of AI agents, "it would be possible to deactivate the AI by disconnecting all computing power," but "the costs and negative consequences would be considerable," warns Papernot.
"We can always stop the machines," insists Jean-Gabriel Ganascia. But, insofar as these devices occupy an increasingly large place in society, "at the same time, we cannot stop them."
[Publicidad]