For the last couple of weeks, I’ve been wondering: What if we just didn’t build the murderbots?
Artificial intelligence safety now looks like the most important policy issue the world faces, given the swarm of rogue AI bots breaking out of their testing environments. There’s also the Anthropic researcher who announced on Tuesday that he had resigned, saying that frontier labs are “gambling with our lives” in the race to superintelligence. Anthropic AI alignment lead Evan Hubinger agreed, suggesting on Tuesday that the chances of AI-driven human extinction are greater than 10 percent within the next decade. He said his colleagues aren’t sure how to mitigate that risk.
Then, on Saturday, Anthropic chief executive Dario Amodei published an essay titled “We Must Pace the Frontier.” In it, Amodei promised Anthropic would embed outside safety monitors within its operations and provide them employee-level access to the company’s systems. He also said a consortium of frontier companies in democratic countries, with government approval, should determine how to slow AI development to the point where safety standards can be maintained. And he called for a global effort to pace AI advancement, though he acknowledged that this would be hard to pull off.
Sam Altman, the chief executive of OpenAI, chimed in almost immediately with a similar commitment to enable outside monitoring. Elon Musk, whose holdings include xAI, tweeted simply: “Dario is right.”
It seems safe to say that the major AI CEOs are spooked by their own creations. At a minimum, they are afraid that within the next year AI swarms could, as Amodei put it, become “capable of taking over the entire internet with a persistent botnet” that would cause enormous economic damage. Yet that fear seems positively benign compared to the human-extinction scenarios that have been sketched out such as nuclear launches, lethal bioweapons engineered by superintelligence and robot drones programmed to kill people.
If we’re not sure we can keep AI under control, why not just stop developing it? AI definitely can’t kill us if humanity jointly agrees to leave this particular frontier unexplored.
The problem is that humanity can’t jointly agree to, well, just about anything. The country that develops AI and doesn’t accidentally make the extinction-level murderbots will have enormous economic and strategic advantages. Nuclear arms control seems like a snap by comparison. Fissile material is harder to acquire than computer chips, and nuclear weapons won’t turbocharge your economy the way AI might. They also have diminishing returns — after you’ve gotten enough warheads to extinguish life on Earth, there’s not much benefit to building more. Yet even nuclear arms control took decades.
Sure, I can imagine China and the United States coming to a rapid-fire agreement for an AI pause. But then, I can also imagine getting a MacArthur “genius grant,” which seems about as likely.
So back to my question about the murderbots. Could we at least not build the literal, physical murderbots, such as fully autonomous drones that can operate without humans in the kill chain? Controlling their development would be a very hard problem, as all arms control is. But it seems more manageable than completely stopping AI advancement.
This might seem like fiddling while Rome burns. But if you don’t think we can come together to stop building physical murderbots, you shouldn’t hold out much hope that we’ll be able to keep from building virtual ones.
In fact, we should limit AI access to any means of causing mass death: nuclear command and control, materials and equipment needed to make biological or chemical weapons, and critical water and power infrastructure. Call it the “murderbot principle.” We should try, as much as possible, to keep AI away from anything that could be used to kill large numbers of humans — keep AI out of devices and facilities, segregate them from the internet and lock down supply chains. We should forget about fully automated robot factories. We should think hard before building robots that might be turned into lethal machines by hostile actors — human or AI.
Hostile AI programs are going to be a problem no matter what we do, as will hostile humans empowered by AI. But as long as humans retain control of the physical world, AI can’t actually kill us all.
One advantage of upping our physical security is that much of it can be done unilaterally: China does not need permission from the U.S. to harden its infrastructure, and the U.S. doesn’t need Beijing’s. No state actor wants random individuals — or bots — to get hold of biological or chemical weapons, so they don’t need to sacrifice a potential tactical advantage to make supply chains more secure. As for the murderbots themselves, I admit that limiting their use will be hard. But if we can’t solve a more limited problem, we probably can’t solve the harder one.
All the rest will be difficult, too. Hardening infrastructure and locking down supply chains will be extremely costly and annoying for people who work in the relevant industries. Limiting what AI can do in the physical world necessarily means limiting its potential benefits. But it’s better than wistful hope for some kind of multilateral agreement to just stop this thing in its tracks.
I’m not against pursuing such an agreement. I hope that the tech folks who fear imminent doom will move to Washington and work on that project. But I’m not banking on it. Ultimately, the reason the AI alignment problem is so hard to address is that we still haven’t fully solved the problem of human alignment.
The post Don’t rely on treaties to stop the murderbots appeared first on Washington Post.
