Researchers at Stanford University have used a generative artificial intelligence model to design entirely new viral genomes from scratch, and 16 of them turned out to be fully functional viruses, a result the team describes as a scientific first. The study was published in the journal Science.
The AI system, built on the "Evo" family of genomic models, was trained on millions of naturally occurring DNA genomes. Once trained, it was tasked with generating brand-new genetic sequences rather than merely analyzing existing ones. According to the BBC, the model produced 302 candidate genome designs, which scientists then synthesized in the laboratory. Of those, 16 successfully assembled into working bacteriophages, viruses that infect bacteria — in this case the common gut bacterium E. coli — rather than humans or animals.
A turning point with two edges
Scientists involved in the project called the result a "very significant turning point" in synthetic biology, according to CNN, noting that AI-designed genomes had never before been shown to produce viable, self-replicating organisms. Because bacteriophages are already used in "phage therapy" targeting antibiotic-resistant bacterial infections that no longer respond to conventional drugs, the researchers say the technique could eventually help design customized phages tailored to specific drug-resistant pathogens.
Forbes reported that the 16 successful designs differed meaningfully from anything found in nature, suggesting the model was not simply recombining existing viral sequences but generating genuinely novel genetic architecture capable of supporting life at the molecular level.
The advance has also triggered warnings from biosecurity specialists. Because the same generative approach could, in principle, be adapted to design more dangerous pathogens rather than benign bacteria-infecting viruses, experts described the dual-use risk as "urgent," calling for stronger oversight of AI models trained on genomic data before the technology matures further. The researchers themselves have acknowledged the tension between the therapeutic promise of the work and the need to prevent its misuse.
No human- or animal-infecting virus was designed or synthesized in the study, and the bacteriophages involved pose no known risk to people. Still, the demonstration that an AI model can move from pattern recognition to the creation of self-assembling, replicating biological entities marks a shift researchers say the scientific community and regulators will need to reckon with quickly.
