Contact The Editor
Articles By This Author
A group of Stanford University scientists posted a paper online in mid-September, describing a feat that could have been plucked from the pages of science fiction: They used artificial intelligence to design new viruses capable of killing bacteria.
In a world where AI keeps creeping in on uniquely human territory by composing sonnets, writing songs or forging friendships, this seemed to be crossing a new Rubicon. Depending on your belief system, AI was doing what evolution, or the Lord, or scientists working with genome-engineering tools aim to do.
"Machines are rethinking what it is to be human, what it is to be alive," said Michael Hecht, a chemistry professor at Princeton University focused on designing novel proteins and artificial genomes. "I find this very unsettling and staggering. They are devising, coming up with novel life forms. Darwin 2.0."
The paper hasn't yet been peer-reviewed, but it is fueling consternation, critiques and think-pieces on what it all means - and what it doesn't. Reactions span the gamut, from "this changes everything" to a scientific shrug. Are machines about to generate novel forms of life, including one that could kill us all? Or is this a powerful new tool - with capabilities that build on what people have been doing for years with more traditional techniques?
Depends who you ask.
The experiment itself wasn't dangerous, and designing "life" is a far heavier lift than the simple phage - a bacteria-infecting virus - that they created. Scientists used "Evo," a generative AI model trained on the genomes of living things. Similar to how other AI large language models are trained on a massive corpus of text, the most advanced version of Evo ingested about 9 trillion letters of DNA from an atlas spanning all domains of life.
Biology has already been revolutionized by AI tools that allow scientists to predict the structure of proteins and generate new ones - the molecules that make life work. This can be good: Scientists are already using the tools to devise new antivenom therapies for snake bites, invent new antibiotics, improve vaccines and break down harmful "forever chemicals." It can also open up new risks, such as creating novel toxins.
But the experiments in the paper went further, building an entire genome: the code for a simple type of virus, called a phage. Most scientists do not consider viruses to be "alive," because outside of a host, they are unable to reproduce. The Stanford team focused on a phage called phi-X174, which injects its genome into E. coli, where it co-opts the cellular machinery to multiply and then kills the cell, spilling out more phages.
Out of about 300 phage genomes the scientists synthesized and tested in dishes full of E. coli, 16 were functional.
"I'm not the least bit surprised by it," said genome pioneer J. Craig Venter, who led an effort to build a bacterial genome from scratch and put it into a cell in 2010. "What current AI is good at is taking tasks that humans can now do and speeding it up a little bit."
But many experts in the field are surprised by it, even if they don't agree about what this paper represents - or what, precisely, it means about where we are going.
"This is a very momentous development," Gregory Kaebnick, a bioethicist at the Hastings Center for Bioethics, said in an email. "It also seems to me to take us another mile down the road in our always-changing relationship to nature, with possibilities that might be welcome but still give me a stomachache. There's tremendous uncertainty in all this."
• When AI and biology intersect
Brian Hie, a chemical engineer at Stanford and the nonprofit Arc Institute, unveiled "Evo" late last year. The aim was to go beyond individual proteins.
"We've previously shown the models can design a single gene at a time, or two genes that interact with each other," Hie said. "But they can also be used to design entire pathways, and in this case, the entire genome of the very simple bacteriophage."
The team chose the bacteria-infecting phage called phi-X174 for medical, practical and historical reasons. In the urgent quest to develop therapies that can treat drug-resistant infections, scientists are already working to reboot phages to kill bacteria. At a brief 5,400 letters long, it was far more manageable than even a bacterium. And phi-X174 was the first DNA-based organism to be sequenced nearly half a century ago.
The technology didn't gin up these life forms on its own accord, but with an expert human team fine-tuning the model, devising design constraints and prompting the program to propose new DNA sequences. The human scientists then ordered the DNA to be synthesized, assembled it and put it in a test tube with E. coli.
The first sign that phages were functional was the cloudiness of the dish. The ones in which phages killed the bacteria were clear. In more detailed tests, some phages multiplied faster than the original strain. A cocktail of AI-designed phages could overcome bacterial resistance.
While some have feared that AI-designed life is on the horizon, Hie said that was not on his agenda at the moment. He is thinking about making parts of organisms that could perform useful functions, such as manufacturing drugs. "We have not yet started or done anything towards making artificial life," Hie said. "It would be quite difficult."
The feat has ignited a debate over what these phages represent. It's the biology equivalent of asking what it means when AI writes a poem in the style of Emily Dickinson. Is AI inventing art or derivatively riffing? Does the distinction matter?
"If somebody came up with a whole new DNA sequence that made a viral particle that killed E. coli, I would be very impressed," Venter said.
Others want to dig deeper into the data, given that biologists use other techniques to modify genes or create life with new functions.
"This is all stuff we could do before generative AI, and the real question is, are we doing it faster, more efficiently, much more novel - what are really the properties that go beyond what we could do before?" said James Fraser, a structural biologist at the University of California at San Francisco.
Others are more impressed.
"I was gobsmacked. I couldn't believe it. I was giddy, as a scientist and a nerd," said Drew Endy, a synthetic biologist at Stanford, who got an early look at the results from Samuel King, the graduate student who led the project. But he pushed back on the notion that AI created life, comparing the AI model to an exquisite musical instrument that can be used by human virtuosos.
"The AI has not done it, but Sam did it - using AI as a beautiful piano," Endy said.
• What could go wrong?
In 1999, Endy got his first DNA synthesizer, a molecular printer capable of spitting out 200,000 letters of DNA in a month. The bottleneck at the time was: what to write? It was like having a printer without a word processor or a language or a dictionary, he recalled.
AI gives scientists the ability to design DNA sequences that don't exist in nature with relative ease. With that possibility comes risk. Endy thinks that in the future, anyone in the world will eventually have the ability to design any number of toxins or dangerous pathogens.
AI vulnerabilities are often discovered and managed after people are already using the technology. Some scientists are trying to put guardrails on the technology now, to reduce the chances it can be used for harm.
Tina Hernandez-Boussard, associate dean of research at Stanford's School of Medicine, co-chaired a conference on responsible AI in 2024 and gave a talk called "What could go wrong?" Hie, who was still working on Evo, asked if she would collaborate to think through safety and ethics issues - for example, by excluding from its training set pathogens that infect multi-celled organisms.
"Brian's approach was: Let's take that out of the training data so they don't have the option to learn these things," Hernandez-Boussard said.
Kevin Esvelt, director of the Sculpting Evolution group at the Massachusetts Institute of Technology Media Lab, said that the viral redesign the AI is doing isn't yet good enough to be a major biosecurity concern.
"I'm not remotely worried today," Esvelt said in an email.
But he added that he hadn't expected the first generatively designed phage for another 18 months, noting that there could be a "concerning trajectory" - where new tools would eventually allow bad actors to create new immune-evading versions of viruses capable of sparking pandemics.
Julian Savulescu, a professor of medical ethics at the National University of Singapore, said that while this phage project could have medical applications, the technology is propelling humanity into "uncharted biological territory - and there is the possibility of creating novel organisms that could wipe out life forms, including human beings."
(COMMENT, BELOW)