Having trouble getting your head around the AI doomsday talk that is suddenly everywhere?
You are not alone.
Some people have been raising the alarm for a long time about the fact that this technology could pose an existential threat. But now everybody’s talking about it. That’s because the tech titans responsible for artificial intelligence are themselves now freaking out publicly, warning that the technology they are building could wreak havoc upon humanity if wielded by the wrong people — or if it develops itself to a point where it can break free from human constraints, and visit misery upon us all on its own. On Wednesday, OpenAI disclosed six new instances in which its AI models did things their human bosses did not intend.
Among the disasters AI might be used to produce, in case you like your nightmares more vivid: massive cyberattacks on financial systems or infrastructure, biological weapons spreading pathogens we cannot combat, AI-powered military weapons getting out of control, blackmail attempts and other manipulations that would force people to do terrible things to themselves or each other.
Those who lead the AI companies could put the brakes on development themselves, but they won’t, because none of them wants to lose the competitive advantage, and the chance to build even bigger fortunes. So they’re asking the government to do it for them.
Let’s pause there for a moment. Company leaders are begging for guardrails that would slow down their work and make it safer: We can’t control ourselves! Stop us, before we kill a bunch of people! Even the megalomaniacal Elon Musk has signed on.
“All of them know that they’re playing a bad game,” said Neil Thompson, a principal research scientist at MIT’s Computer Science and Artificial Intelligence Laboratory. “This is a game where they know that if everybody continues to race as fast as they can, we’re going to continue to accumulate these dangerous capabilities in a way that is a problem. But they also know that unilaterally they can’t change that game.”
Right now, the profiteering president and the lemmings in Congress are refusing to lift a finger. And they’re not exactly fans of the kind of international cooperation that would be required to get a handle on this technology globally.
Ain’t late-stage capitalism grand?
Some observers have suggested that the AI makers’ public freakouts are actually marketing ploys — look how powerful our technology is, aren’t we amazing! — or attempts to insulate themselves against lawsuits, or efforts to distract us from the very real environmental damage their data centers are unleashing all over the country. But none of those motivations lower the risk here.
“I do not think this is a marketing play to get attention,” Thompson said. “This is a well-founded concern that I and many share about the incredible capabilities that this tool has, but how that also carries with it great risks.”
It’s easy to hear this assessment and just move on.
“It’s a real challenge for people to be able to absorb this kind of risk,” Thompson said.
But sit with this, if you can bear it: Earlier this year, MIT FutureTech, a research group Thompson leads, surveyed 272 AI experts around the world to assess the likelihood and severity of the risks posed by the technology if it remains unchecked. The experts judged a more than 10 percent chance of catastrophic outcomes from AI-powered cyberattacks, creation of AI-enabled weapons, and other scenarios resulting in more than 1 million deaths or more than $100 billion in losses in the next five years.
To most of us, that seems unreal, impossible. And for good reason.
“People are quite right in general to think that most technological changes are not going to change everything overnight,” Thompson said. But no other technology we’ve developed has advanced this fast, and with such broad reach, affecting almost every part of our lives, he said. Normalcy bias makes us resistant to accepting the growing dangers — and will make folks more vulnerable to AI-powered, super-sophisticated financial scams and influence operations. That bias will also mean voters will take longer to see the threat clearly enough to demand that their leaders do more to regulate AI development, and protect citizens from its destructive capabilities.
So we are probably going to have to do this the harder way.
“I think we are likely to see an increasing number of incidents coming out,” Thompson said, meaning attacks by AI-assisted villains, or AI acting on its own in unintended ways. No one wants that to happen, but “I think that will focus public attention both here and elsewhere in the world.”
I guess we have that to look forward to.
“I just hope that consensus happens fast enough that we have as little damage as possible on the road to it,” Thompson said.
Globe columnist Yvonne Abraham can be reached at yvonne.abraham@globe.com.