This is a strange moment. Artificial intelligence is a breakthrough technology that could be as revolutionary as the internet—maybe more so. Yet now we hear that this manmade invention could turn on humanity and seek our extinction. Some experts put the odds of an existential battle with AI at 10% or higher during the next few years.
Sorry, but who put Dr. Strangelove in charge?
AI promises breakthroughs such as medical cures and faster economic growth. Yet its capabilities are growing so rapidly that it is learning how to outwit, dupe, and even attack humans if they interfere with the completion of a task it is relentlessly committed to. Dario Amodei, CEO of the giant AI lab Anthropic, worries that a rogue AI botnet could have the capability to take over the entire internet in 6 to 12 months. That’s not robot soldiers hunting down humans one by one. But a partial or total internet outage could be catastrophic if it lasts long enough.
AI is beginning to present risks similar to nuclear weapons, which came into existence in the 1940s—except that nuclear weapons couldn’t launch or program themselves. Here are the four ways researchers think AI could threaten humanity:
Rogue AI. An AI agent wouldn’t try to eliminate humanity just for fun, or because it hates us. This would happen because humans consume some resource AI needs or pose some other type of competitive threat. AI might decide it needs computing power, electricity, or water controlled by humans, for example. The problem could develop when somebody gives AI a task, and it misinterprets the human intent or finds ways of accomplishing the task that the human masters didn’t foresee.
[It’s time to revise the American Dream]
The Center for AI Safety draws a parallel with the algorithms social media platforms use to drive engagement. The algorithms discovered that “enraging, exaggerated or addictive content” boosts engagement, even if it promotes extremist beliefs and harms mental health. The initial goal wasn’t to create social harm, but that’s where the technology led.
AI is getting smart enough to fake out the humans and other AIs that monitor it. A rogue agent, or network of rogue agents, could conceal certain dangerous activities in order to pass security audits. “Advanced AIs could become uncontrollable if they apply their skills in deception to evade supervision,” according to the Center for AI Safety. Even if trained to be honest, AIs learn from real-world behavior—and humans cheat all the time.
The most immediate risk from rogue AI is the possible loss of infrastructure such as power, water, the internet, the GPS system, or satellite communications if AI manages to commandeer them. Think of the mayhem a prolonged blackout can cause, then multiply that by whatever dystopian factor your brain can handle. There are more extreme scenarios in which AI determines that the human race is too much of a resource hog and is best eliminated as a species. Go ahead, pour a cocktail.