AI just created viruses from scratch

TL;DR: For the first time ever, scientists say they’ve taught AI to create brand-new viruses. The breakthrough could offer a new route for treating infections that antibiotics can’t touch, but you can probably guess the downsides…

What happened: In a new study published in Science yesterday, researchers at Stanford University and the Arc Institute showed that they used AI to write new viral genomes—first by teaching it to recognize the DNA structure of natural viruses, then asking it for blueprints that don’t exist. The team’s AI models, Evo 1 and Evo 2, trained on a massive library of genetic data, then specifically on viruses related to Phi X-174, a natural virus that only infects the E. coli bacteria.

Of the 700,000 potential viruses generated by Evo, nearly 300 were built in the lab, and 16 turned out to be viable. (Some multiplied even faster than Phi X-174.)

Why it matters: A mixture of the AI-designed viruses overcame E. coli strains that had learned to fend off Phi X-174 (which a comparable mix of natural viruses couldn’t do). It’s an early proof of concept that AI could be used to treat drug-resistant bacterial infections. It’s also a turning point for AI use in biology, as models move from analyzing and predicting genomes to writing complete, functional ones.

Voluntary safety: As anyone who saw the Bond movie No Time to Die knows, engineered bioweapon plots don’t end well. And right now, the only thing keeping that scenario fictional is the honor system. The researchers in the study wanted to be “extra careful” about the potential of generating genomes of viruses that could harm humans, and deliberately excluded viruses that could infect humans, animals, and plants from the training data.

Health security scientists at Johns Hopkins, who weren’t involved in the study, commended the team on the precautionary step. As they put it, “The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.”

Bottom line: The models are a step toward new treatments, but they’re also a preview of the next AI safety fight. For now, the only guardrail is what a research team decides not to teach its model. —LC