In the “Terminator” movie series, the artificial-intelligence system known as Skynet is given control of the nation’s nuclear weapons and uses it to try to destroy humanity.
One reason that apocalyptic scenarios involving AI are implausible is that, having seen how that particular movie ends, we are unlikely to be foolish enough to give AI custody of the most destructive weapons on the planet.
The unthinkable is nonetheless a hot topic of conversation, given that a researcher at Anthropic just resigned and issued a doomsday warning.
According to Jacob Coxon, the rush to develop AI is “gambling with our lives,” since it will soon create “superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.”
A current Anthropic employee then chimed in: “We really do earnestly believe AI could kill all humans!” He put the chances at greater than 10 percent in the next decade.
The idea is that Anthropic and similar firms are, in effect, running their own Manhattan Projects, with consequences unknown and potentially very dire.
Now, ChatGPT as we know it doesn’t seem particularly well-suited to destroying humanity. It is a souped-up version of Google that is impressively fast, but also mistake-prone. If it can mess up basic math or the New York Yankees roster, surely it can’t carry out the complex tasks necessary to fashion an apocalyptic event.
Tech companies, though, have labs working on so-called frontier models that are much more powerful and reportedly developing more quickly than expected. The worry is that an AI model itself will begin to develop new AI very quickly in an “intelligence explosion,” the equivalent of Skynet becoming self-aware at 2:14 a.m. Eastern Time on Aug. 29, 1997.
The thinking is that if such an AI model goes rogue and turns against humanity, it will be too clever and fast for us to combat.
Yet it’s hard to see the mechanism that such a hostile superintelligence would use to commit omnicide. The rogue AI would presumably have to do something in the real world — say, launching nukes, manufacturing killer robots or releasing a deadly virus.
If it lacks physical agency, AI could presumably be unplugged or contained. (It was after humans tried to deactivate Skynet, by the way, that the system decided we were a threat that had to be neutralized.)
Some doomsayers posit a malicious person using advanced AI to come up with a design for a killer virus and a plan to release it without detection. This is a different scenario, though, from AI turning against people of its own volition. In a world years from now when much more advanced AI is a reality, we will need to think through safeguards to keep dangerous systems — nukes, biolabs — from being compromised by AI-enabled cyberattacks or exploited by criminals and terrorists.
Science fiction is fascinating and sometimes prescient. Works by the likes of Philip K. Dick, Aldous Huxley and Arthur C. Clarke had genuine insights into the future. But that doesn’t mean we are living in the plot of “The Terminator,” and we shouldn’t scare ourselves into forswearing transformative technology.
Rich Lowry is on X
@RichLowry.
