Scientists at Stanford University and the Arc Institute have crossed a line that, until this week, existed only as a possibility. They used artificial intelligence to design complete viral genomes from scratch, then brought those digital blueprints into the laboratory and watched some become functioning organisms.
The peer-reviewed result, published in Science, moves the discussion beyond what AI might someday do in biology. These viruses were bacteriophages, pathogens that infect and kill bacteria rather than people. However, the experiment still puts a new capability in plain view: a computer-generated genome can pass a physical, real-world test.
That gives the work two immediate dimensions. It points toward a possible new tool against difficult bacterial infections while raising a biosecurity question researchers and specialists are already debating as the science moves faster than the frameworks meant to govern it.
The breakthrough is the real-world test: an AI-generated DNA blueprint became a functioning virus in the lab.
What Actually Happened in the Lab
Most of the designs failed, and that makes the sixteen successes far more revealing on close inspection.
The team generated thousands of candidate genome designs with AI, then chose 302 for laboratory synthesis. Those DNA sequences were chemically manufactured and introduced into bacteria so researchers could see which designs actually functioned.
Only sixteen of the 302 produced viable bacteriophages capable of replicating and killing E. coli. Brian Hie, an assistant professor at Stanford, described the result to the BBC as “a next step in the complexity that’s designable by generative AI… the first time generative AI has been used to design a complete genome, it’s something that can replicate and have other functions inside cells… this was new territory for us.”
The model generated possibilities; laboratory testing decided which ones were real. Laboratory performance determined the sixteen successes.
The Models That Wrote the Blueprints
The experiment relied on two systems, Evo 1 and Evo 2, trained to recognize the architecture of genetic information across an enormous biological record.
Reports note that both models were trained on millions of genomes drawn from across life forms. The goal was to learn patterns such as gene arrangement, conserved sequences, and the biological constraints that separate a working genome from a meaningless string of genetic letters.
That training allowed the systems to generate novel sequences within a bacterial-virus framework, including sequences no pathogen in nature has ever carried. Arc Institute notes that its Evo models had already generated individual proteins and multi-component biological systems, while a complete, functional genome remained a longer-standing goal.
To ensure safety, the researchers strictly excluded genetic data from viruses capable of causing human disease from the training. That deliberate biosecurity guardrail influenced what the models could propose in this experiment.
The Medical Problem Behind the Science
The reason for designing new bacteriophages sits inside a much older problem: antibiotic resistance is steadily shrinking the treatment options available to clinicians worldwide.
Bacteriophages, or phages, are viruses that target and kill specific bacteria. Scientists have studied them as possible alternatives to antibiotics for more than a century, and they remained in limited clinical use in parts of Eastern Europe, Russia, Georgia, and Poland after antibiotics became the dominant Western treatment in the 1940s.
Interest in phage therapy has returned because bacteria can evolve to survive drugs. Stanford and the Arc Institute are testing whether AI might widen the pool of phages researchers can explore.
In the study, a mixture of AI-generated phages overcame resistance in E. coli strains that a naturally occurring phage mixture could not clear. The finding involved two bacterial strains in a laboratory, not a treatment or clinical result, but it gives researchers a concrete reason to keep investigating.
Save this article
Enter your email address and we'll send it straight to your inbox.
Why These Particular Viruses Made Sense as a First Test
For a first attempt at whole-genome AI design, the researchers chose a small virus whose biology was already extensively studied: the bacteriophage Phi X-174.
Phi X-174 has a compact genome mapped in fine detail, making it a comparatively manageable target for whole-genome AI design. Stanford-associated researchers had already shown in 2012 that its overlapping genes could be separated, altered, and re-synthesized while still producing infectious phages.
That earlier work was done by hand, without AI, and established that this genome could respond predictably to modification. The new experiment asked a different question: could AI generate entirely novel genomes within the same bacterial-virus framework?
Most synthesized candidates failed to produce viable viruses. Some, however, replicated and killed bacteria, turning computer-generated sequences into a physical laboratory result.
The Safety Question That Came With the Results
The same experiment that demonstrated a new biological capability also sharpened a problem science policy has not fully resolved: who sets the rules once AI can design a viral genome?
Thomas Inglesby and Moritz Hanke of Johns Hopkins reportedly put the concern plainly: “the ability to design a viral genome with AI now exists, and the rules for doing it safely do not.” The Stanford and Arc Institute team restricted its work to bacteriophages. Still, experts note that the training-data exclusions used here are not fixed limits on what similarly capable future models could learn or what actors with different intentions might attempt.
Filippa Lentzos of King’s College London pointed to practical controls. The argument reflected across coverage of the research is that a critical safeguard may sit at the point where a digital genome design becomes physical DNA through a synthesis provider.
That step occurs before any laboratory testing begins, placing it directly inside the debate over oversight.
What This Does and Does Not Mean
The result is significant, but its boundaries matter: no AI-designed therapy is available to patients, and no human-infecting virus was created.
What the study demonstrates is very striking. Under specific laboratory conditions, after extensive screening and testing, AI-generated whole-genome designs can become working bacteriophages that replicate and kill bacteria.
Any path from this result to a therapy for patients would still require animal studies, trials for humans, safety evaluation, and better regulatory review. None of those stages has been bypassed, and the distance between this experiment and a clinical treatment remains measured in years of additional work.
The study does establish that AI can now participate in generating biological candidates at a scale natural discovery cannot match. The governance questions now sit beside the scientific ones: who can access these tools, what sequences can be ordered, and who is checking?
As AI tools become capable of designing functional genomes at scale, should the responsibility for oversight fall with the researchers, the AI developers, the DNA synthesis companies, or government regulators?
