The real AI threat to humanity is a super-virus created in a garage

This is Everyday Science with Clare Wilson, a subscriber-only newsletter from The i Paper. If you’d like to get this direct to your inbox every week, you can sign up here. Hello, and welcome back to Everyday Science. You may have heard the alarming claim that there’s more than a 10 per cent chance humanity will be wiped out by super-intelligent AI. This assertion, made by a researcher at Anthropic, was scant on detail about exactly how the technology could kill us all. But this week I heard about one possible way, which wasn’t previously top of my list of things to worry about: AI could be used to make a lethal pathogen and trigger another pandemic. Shorts The warnings came at a press conference held this week by disease experts from Imperial College London’s biosecurity network. “I believe it’s highly unlikely we’ll get an extinction-level event,” said epidemiologist Professor Neil Ferguson. “That doesn’t mean I think there’s zero threat.” So, what might such an AI-created bioweapon look like and how could we stop it from being created? Starting template for bioweapon One reason this threat has to be taken seriously is that pandemics with high fatality rates have previously arisen naturally without needing any help from super-intelligent AI. The Spanish flu, which hit at the end of the First World War, killed an estimated 5 per cent of the world’s population. In more recent memory, the Covid pandemic killed a smaller fraction of people – perhaps about 0.2 per cent – but was still hugely damaging and frightening. The UK Health Security Agency last year made a list of 24 groups of pathogens that have most potential for triggering a dangerous pandemic. It includes members of the coronavirus family, to which Covid and Sars belong, as well as others such as mpox, bird flu and Ebola. The agency drew up this list so researchers can focus on developing vaccines and treatments in case of natural pandemics – but they could also be the starting template for an AI-designed super-bug. AI changes things because the genetic engineering of viruses or bacteria used to need a lot of money, large well-equipped laboratories and a great deal of expert knowledge. In other words, it would only have been possible in universities or other state-funded laboratories. ‘Garage biology’ No matter how much leaders of states such as Russia or Iran may dislike the West, it is hard to imagine them releasing a deadly virus, because it would also kill their own citizens. We have seen with Covid how infectious diseases have no respect for borders. But in the past few decades it has become more possible for anyone to dabble in genetic engineering of viruses or bacteria, in the so-called “garage biology” hobbyist movement using cheap second-hand lab equipment. You can now even buy a self-contained kit that lets you insert a gene from jellyfish into harmless bacteria, and watch your genetically-engineered colonies of bugs glow in the dark as they multiply. The concern is that AI could now provide instructions on how to make more dangerous bugs. “One fear is that AI allows people to put together the instruction manual more readily. That is something that governments and Imperial are thinking about,” said Professor Wendy Barclay, a virologist at Imperial College London. How close is AI to designing a deadly virus? The most advanced generative AI programmes have certainly proved capable of some impressive feats, such as solving previously insurmountable maths problems, including finding a solution last week to one called the Navier-Stokes problem concerning the movement of fluids. In some ways, biology is harder for AI to crack than maths or physics, because it is more complex. But in 2020, Google announced its AI programme had “largely solved” a 50-year-old biology problem of predicting the three-dimensional structure of proteins, based only on their sequence of amino acid building blocks. High precautions are taken around chickens with bird flu – which could cause the next pandemic (Photo: Peter Garrard Beck/Getty) However, it is a much trickier problem to go from a pathogen’s genetic sequence to knowing how it will behave in its human hosts – in other words, whether a DNA tweak would make something more deadly. They are not there yet. Some university groups are currently using AI for benign purposes to design new genes and even whole new viruses. But the computers rarely get it right first time: each attempt has to be tested on cells in a dish or animals before it’s perfected, said Professor Tom Ellis, co-director of Imperial’s Centre for Engineering Biology. “A lot of what it would design at the DNA level will not work first time,” he said. “And the experiments to validate and test [the DNA sequences] are still long and arduous.” Those experiments are not what someone working out of their garage could achieve – not to mention they would also risk dying from their own bioweapon without high-quality containment equipment. But one hallmark of AI is how fast it is advancing. So, it would be irresponsible not to start considering how individuals could be stopped from carrying out such a plan in future, said epidemiologist Professor Azra Ghani. Security checks At the moment, firms making the raw materials that could be used to make bioweapons – like certain viruses or DNA that encodes potentially dangerous genes – tend to have multiple security checks. These ensure the materials are only sent out to scientists working at regulated laboratories, who can show they are studying potential treatments, vaccines and so on. But these are currently voluntary. What we need are international agreements to make such security checks mandatory, said Barclay, the virologist. “We need some level of regulation of access to the tools.” Stricter regulation will be on the agenda at a series of meetings on pandemic preparedness around the United Nations General Assembly in New York next week. “The aim is to think ahead of time about how these threats are changing and how we’re going to detect novel threats,” said Ghani, who will attend the meetings. Ultimately, Ghani said, most scientists think a future pandemic is more likely to have natural origins than to arise from an engineered bioweapon. And our best chance of combating these natural pandemics with new drugs, vaccines and diagnostic tests? That could also be help from AI. I’ve also written It is claimed that cherry juice and supplements are the perfect post-workout recovery aid, but what does the science say? I’ve been watching Two months after it opened I have finally seen Christopher Nolan’s The Odyssey, and it was worth the wait. As someone who grew up reading the Greek myths, I found this modern retelling true to their spirit in terms of excitement, casual cruelty and gut-churning horror. The Cyclops in particular was not for the faint-hearted. And nor is the film for the weak-bladdered, at nearly three hours long. This is one movie trip not to be accompanied with a soft drink.