There’s been some discussion lately about the danger of AI in biology. One of the sparks was a paper in Science, in which King et al. showed that genome language models can be used to design functional bacteriophages - viruses that infect bacteria. The paper raised safety concerns, which were met with some ridicule and reassurance from experts in different fields. For example, Noah Smith’s post raised the concern that AI could make designing dangerous viruses accessible even to non-experts, with disastrous consequences. Claus Wilke’s response, “I’m sorry, you’re not going to die from an AI-engineered supervirus,” highlighted the gaps between what AI can currently do and what Smith’s scenario would require. Those gaps include both biological design and the practical challenge of turning a design into a working biological agent. I agree that these are serious barriers but I also think an important part of the concern remains unanswered.
The question is whether AI makes dangerous biological work substantially easier than it was before. Does it make existing knowledge more usable? Does it allow a smaller group of people to accomplish something that previously required a much larger team? If the answer is yes, that deserves a serious discussion, even if completely AI-designed superviruses are just buzzwords at this point in time.
We already know that existing biological agents can cause devastating harm. Anthrax is one example. Although it is treatable, inhalation anthrax is almost always fatal without treatment, and even with aggressive treatment, only about 55% of patients survive, according to the CDC. In 1979, an accidental release from a Soviet military biological facility in Sverdlovsk killed dozens of people. The investigation showed that existing pathogens could have terrible consequences without any help from AI or even sophisticated molecular biology.
Anthrax doesn’t spread from person to person, so that example is rather limited in scope, but Soviet bioweapon programs went beyond using pathogens as they occur in nature. According to former Soviet biological weapons scientist Sergei Popov who defected to the west in the 1990s, researchers worked on making bacteria resistant to multiple antibiotics. The intention was to make existing diseases harder to treat, and this work was happening decades before today’s AI tools. Popov’s interview with PBS
The Soviets also pursued biological weapons involving contagious diseases, so their ambitions were not limited to harming people exposed during the initial attack. That is different from demonstrating that they successfully made bacteria spread more easily between people, and we should be careful not to confuse the two. But the broader concern remains: efforts to turn existing pathogens into more dangerous weapons have a long history.
None of this tells us that a small group with AI could reproduce what a massive state program attempted. It does explain why the difficulty of designing a completely artificial virus is only part of the discussion. The question is how much AI could lower the barriers to dangerous work based on biology we already know.
The availability of treatment is also only part of the story. A medicine being effective does not mean everyone who needs it will receive it in time. A large number of people becoming seriously ill at once can overwhelm hospitals, exhaust supplies, and disrupt the distribution of care. For contagious diseases, continued transmission can make that pressure worse.
None of this means a 100% fatality rate, and overwhelmed hospitals do not automatically mean the collapse of civilization. But the consequences can still be catastrophic. Our ability to withstand such an event depends on medical teams, functioning supply chains, and emergency services that can keep operating under pressure. “We have a treatment” is reassuring only to the extent that we can actually deliver it.
So what does AI have to do with this?
Turning biological knowledge into a weapon is a serious undertaking, which is barred by signficant resource constraints. It requires expertise across different fields, practical experience, equipment, and resources. Very few people are well versed in all of it. Having access to scientific information is not the same as knowing how to make it work, and AI does not make those difficulties disappear.
Still, making unfamiliar technical knowledge easier to understand and use is one of the things AI appears increasingly useful for. The concern is that it could reduce some of the barriers that currently keep dangerous biological work beyond the reach of small groups. How much it can reduce those barriers in practice is the question we need to answer.
There’s a temptation to point to people running businesses or building software with teams of AI assistants and assume that biology will follow the same path. That analogy only goes so far. Experiments fail, practical skills matter, and a convincing answer on a screen can be completely wrong. We need evidence that AI assistance changes what people can actually accomplish in biological work.
But those limitations should not end the discussion either. AI would not have to replace every expert or solve every problem to make a difference. Even a partial reduction in the expertise required could matter to bridge missing expertise. Groups like Al-Qaeda who have shown interest in the past in developing bioweapons, and have arguably recruited former students who went to elite universities in the west to execute their sinister plans, could now have a much lower barrier-of-entry to this dangerous game.
My concern, then, is that the debate is spending too much time on whether AI can invent a civilization-ending supervirus. We should also be asking whether it can make existing biological knowledge easier for a small group with malicious intentions to use. We do not yet know how far that capability extends, and we should be careful about claiming otherwise. But we do not need certainty about the end of civilization to take the risk seriously.
A lot of what I know about the potential harm of bioweapons comes from two main sources, for those who want to deepen their knowledge
Spillover by David Quammen, describes cases of zoonotic jumps of pathogens from animals to humans. This book was written much before Covid and made me think how lucky we are that such pandemics haven’t occurred before
Biological War: A Scenario by Annie Jacobsen, discusses a scenario in which an extremely aggressive variant of Y.pestis is released by accident. This variant is not only resistant to all antibiotics we have, but also harbors virus particles that are released when the bacteria is eliminated.