For the first time, artificial intelligence has been used to generate 16 brand-new, previously undetected viruses. These new synthesized viral genomes are unique, and instead of infecting humans, they infect illness-causing bacteria itself. For an explanation on how this works and the implications for the future of medicine, GBH Morning Edition host Tori Bedford spoke with MIT biologist and associate professor Kevin Esvelt. What follows is a lightly edited transcript.
Bedford: When we think about AI inventing new viruses, this sounds really scary, but the idea here is that these could actually help us, right? Can you explain how this works?
Esvelt: The notion here is that there’s a lot of bacteria that are becoming resistant to antibiotics, and some folks have tried to use viruses to kill these bacteria instead. The problem is that that requires us to go out and find a lot of viruses in nature, and because they’re already present, they already exist, many bacteria have resistance to the viruses too. So the idea here was, what about using AI to generate new variants of viruses that could overcome resistance? If you get an AI that is good enough at making new viral variants, you could always come up with a new variant that would overcome the resistance of any given bacterial infection.
Bedford: What does it mean that AI created these? How does AI create a virus?
Esvelt: Yeah, so a lot of people are saying, “Oh, these are completely new viruses.” They’re actually not. They’re more than 90% identical to viruses of a particular family that nature created already. What’s different about them is simply that they were designed on a computer, and they’re about as different from the ones in nature as different viral variants of an existing viral family are from each other. That is, nature has only explored a tiny fraction of all the possible viruses that could exist, even within one kind of virus. Here, the AI is just filling in the blanks, creating the new ones that nature didn’t create.
Bedford: OK. And how would this be used?
Esvelt: Well, this touches back on this question of how best can we use viruses against human infections, because we have a regulatory system that normally approves “one drug for one bug,” as we say. You get infected with a bacterium, say a Staphylococcus infection, and so normally you get approved to treat Staph with one drug. But here, the AI can generate a lot of new viruses. So it’s kind of an open question: do you want to just generate these, say, 16 new viruses? Could you put them together in a bundle and ask the FDA to approve those? This has always been the problem with treating bacterial infections with viruses. There’s lots of viruses, and it works best if you deliver them as a group rather than any single virus.
Bedford: And then with the creation of these viral genomes, there is concern about how the technology could be misused, right? Can you talk about that?
Esvelt: Yeah, so this was done extremely safely. There is no way that a virus that infects bacteria can infect us. We’re just wildly different. But at the end of the day, a method that can generate viral variants of bacteria-infecting viruses could be applied, in principle, to any other kind of virus. So the researchers who did this study were very cautious about this. They deliberately did not train their AI on any viruses that infect pretty much anything multicellular. They wanted to make it as hard as possible for their technology to make variants of human viruses. The problem is, once it’s been done, it’s been done. Someone could train a new AI that would work on human viruses. This is the problem when you make the AI what they call “open weights.” You essentially share the brain of the AI with everyone so that anyone can train it to do anything further. There’s really not much stopping people from training it on human viruses. And the downside of this is, of course, pretty much any virus that we currently keep in check with a vaccine, there is some variant that could overcome our immunity and then spread. So the unsettling conclusion is that this study is one more step in the road towards being able to make unlimited pandemic viral variants in humans.
Bedford: So essentially creating a biological weapon, like a mass bioweapon?
Esvelt: Even worse, it is a biological weapon, but it would be a pandemic weapon. And you might say, oh well, SARS-2 was not that bad, right? I mean, it killed a million Americans, so it actually was that bad. But you might say Omicron wasn’t that bad if you’ve already had COVID. Well, it still killed an awful lot of people. And if you completely evade existing immunity, then the infection is going to be about as bad as it was originally in a lot of cases. So we should begin thinking now how we’re going to prepare for a world where people can just cause pandemics at will.
Bedford: That’s terrifying. OK, so it opens up a lot of exciting new possibilities to fight viruses, but also could create the next global pandemic and biological weapon that could wipe out a ton of people is what you’re saying.
Esvelt: It’s hard to see how science will not learn how to do that and it’s hard to see how on our current trajectory, where people want to make tools to understand how to make better vaccines that natural viruses can’t evolve around, how do you do that without learning how to make a virus that can escape? The AI tool is dual use that way. So it’s possible that you will not be able to do this in the United States. It’s hard to see how we can keep anyone from doing it anywhere. And right now, you can’t use federal funds to make a virus that could infect humans that might escape our immunity. And that actually is as of two weeks ago. It was actually permissible before that. It was scrutinized, but it was permissible. As of now, that is banned, but there’s an exception for computational advances. So you can train the AI just fine. You just can’t actually make the viruses to test whether it worked. But there’s a loophole. You could do it in mice. There’s nothing stopping someone from training the AI and then generating a variant of a mouse virus and seeing whether it works. So one thing the government should consider doing is closing that loophole including computational methods of designing pandemic viruses in the ban.
Sign up for GBH newsletters
