If AI goes rogue, there are ways to fight back. None of them are good.

It is Advice is as old as tech support. If your computer is doing something you don't like, try turning it off and on again. When it comes to the growing concerns that a sophisticated Artificial intelligence system could go rogue so catastrophically that it could pose a risk to society, or even humanityIt's tempting to fall back on this kind of thinking. An AI is simply a computer system designed by humans. If it's not working properly, can't we just turn it off?

In the worst case scenario, probably not. This is not just because a sophisticated AI system could have a self-preservation instinct and resort to desperate measures to save itself. (Versions of Anthropic's large language model Claude resorted to “blackmail”. This is also because the fraudulent AI may be too widespread to be taken down. Current models like Claude and ChatGPT already run in multiple data centers rather than on one computer at one location. If a hypothetical rogue AI wanted to avoid being shut down, it would quickly copy itself across the servers it has access to, thus preventing unfortunate and slow humans from pulling the plug.

In other words, killing a rogue AI may require destroying the Internet or large parts of it. And that is no small challenge.

This is the challenge that preoccupies Michael Vermeer, a senior scientist at the Rand Corporation, the California-based think tank once known for its pioneering work on nuclear war strategy. Vermeer's most recent research addressed the potential catastrophic risks hyperintelligent AI and told Vox that when these scenarios are considered, “people dismiss these wild options as viable possibilities” of how people might respond without thinking about how effective they would be or whether they would create as many problems as they solve. “Could we actually do this?” he asked himself.

In one Current paperVermeer looked at three of the most commonly suggested options by experts for responding to what he calls a “catastrophic loss of control in AI.” He describes this as a rogue AI that locked people out of key security systems and created a situation “that threatens the continuity of government and human well-being to such an extent that the threat would require extreme measures that could cause significant collateral damage.” Think of it as the digital equivalent of that The Russians are letting Moscow burn to repel Napoleon's invasion. In some of the more extreme scenarios Vermeer and his colleagues imagined that it might be worth destroying a large portion of the digital world to kill the rogue systems within it.

In (probably) ascending order of potential collateral damage, these scenarios include the use of another specialized AI to counter the rogue AI; “Shutdown” of large parts of the Internet; and detonating an atomic bomb in space to create an electromagnetic pulse.

None of these options leave you with a particularly good feeling after completing the work.

Option 1: Use an AI to kill the AI

Vermeer envisions creating “digital vermin,” self-modifying digital organisms that would colonize networks and compete with rogue AI for computing resources. Another possibility is a so-called hunter-killer AI, which is intended to disrupt and destroy the enemy program.

The obvious downside is that when the new killer AI becomes advanced enough to have any hope of completing its mission, it may itself go rogue. Or the original rogue AI could exploit it for their own purposes. By the time we actually consider options like these, we may be past the point of caring, but the potential for unintended consequences is high.

Humans don't have a particularly good track record of introducing one pest to eradicate another. Think about it the cane toads The beetle was introduced to Australia in the 1930s and has never done much to eradicate the beetles they were supposed to eat, but has killed many other species and continues to cause devastating environmental impacts to this day.

However, the advantage of this strategy over the others is that it does not require the destruction of actual human infrastructure.

Vermeer's article considers several options for shutting down large parts of the global internet to prevent the spread of AI. This could involve manipulation of some of the fundamental systems that enable the Internet to function. One of these is Border Gateway Protocols, or BGP, the mechanism that enables the exchange of information between the many autonomous networks that make up the Internet. A BGP error was something caused a massive Facebook outage in 2021. In theory, BGP could be used to prevent networks from communicating with each other and bring down parts of the global Internet. However, due to the decentralized nature of the network, this would be difficult and time-consuming to accomplish.

There is also the “Domain Name System” (DNS), which translates human-readable domain names like Vox.com into machine-readable IP addresses and is based on 13 servers distributed worldwide. If these servers were compromised, it could block access to websites for users around the world and potentially our fraudulent AI. However, it would be difficult to shut down all servers quickly enough to prevent the AI ​​from taking countermeasures.

The paper also considers the possibility of destroying the physical infrastructure of the Internet, such as: Subsea cables, through which 97 percent of the world's supply comes Internet traffic travels. This has recently become a problem in the world of national security between people. Internet service was disrupted due to suspected cable sabotage Islands around Taiwan and on islands in the Arctic.

But there are simply too many cables and too many redundancies built in around the world for a shutdown to be feasible. That's a good thing if you're worried that World War III could shut down the global internet, but a bad thing if you're dealing with an AI that threatens humanity.

Option 3: Death from above

In one 1962 test known as Starfish PrimeThe United States detonated a 1.45 megaton hydrogen bomb 250 miles above the Pacific Ocean. The explosion caused an electromagnetic pulse (EMP) so powerful that it knocked out streetlights and telephone service in Hawaii, more than 1,600 miles away. An EMP creates a voltage surge strong enough to destroy a variety of electronic devices. The potential impact in today's much more electronically dependent world would be much more dramatic than in the 1960s.

Some politicians, like former House Speaker Newt Gingrich, have done so spent years Warning of the potential damage an EMP attack could cause. Thanks to US intelligence, the issue was back in the news last year Development of a nuclear device for launch into space.

Vermeer's article assumes that the United States is intentionally detonating warheads in space to cripple ground-based telecommunications, energy and computer infrastructure. It could take an estimated 50 to 100 detonations total to cover the U.S. landmass with a sufficiently powerful pulse.

This is the ultimate blunt tool where you want to be sure the cure isn't worse than the disease. The effects of an EMP on modern electronics – which may include surge protection measures in their design or could be protected by buildings – are not precisely known. And in the event that AI survives, it wouldn't be ideal if humans had disabled their own energy and communications systems. There is also the alarming prospect that if other countries' systems are affected, they could retaliate against a nuclear attack, no matter how altruistic the motives.

Given the unattractiveness of these individual approaches, Vermeer is concerned about the lack of planning he sees among governments around the world for these scenarios. But he points out that it is only recently that AI models have become so intelligent that policymakers have begun to take their risks seriously. He refers to “smaller Instances from Loss the control of powerful systems, which I think should make it clear to some decision makers that we need to prepare for this.”

In an email to Vox, AI researcher Nate Soares, co-author of the bestselling and nightmare-inducing polemic, said: If someone builds it, everyone diessaid he was “encouraged to see elements of the national security apparatus beginning to grapple with these thorny issues” and largely agreed with the article's conclusions – but was even more skeptical about the feasibility of using AI as a tool to control AI.

Vermeer, for his part, assumes extinction An AI disaster is a low probability event. But loss of control scenarios are likely enough that we should be prepared. For him, the conclusion from the paper is that “we are not prepared for the extreme case of globally distributed, malicious AI. We are left with only poor options.”

Of course, we also have to take into account the old military maxim that the enemy gets a voice in every question of strategy. These scenarios all assume that humans should retain basic operational control over government and military command and control systems in such a situation. Since I recently reported for VoxThere are reasons to be concerned about the introduction of AI into our nuclear systems, but AI actually launching a nuclear bomb is probably not one of them, at least for now.

Still, we may not be the only ones planning ahead. If we know how bad the available options would be for us in this scenario, the AI ​​will probably know that too.

This story was produced in collaboration with Outrider Foundation And Partner to promote journalism.