A genome is a string of nucleotides, and the same kind of artificial intelligence that generates prose can, in principle, generate one. Researchers at Stanford University and the Arc Institute have now done exactly that. They used genome language models, AI systems trained on billions of nucleotides the way chatbots are trained on words, to write complete bacteriophage genomes from scratch, synthesized nearly 300 of them, and found 16 that produced working viruses. The results appear in Science alongside a companion Perspective warning that the governance needed to steer this capability safely does not yet exist.
Bacteriophages, viruses that infect bacteria, are a convenient test bed for genome design. The team chose as their template phi-X174, a tiny phage of just 5,386 nucleotides and 11 genes that infects E. coli. It was the first genome ever sequenced, by Fred Sanger’s group in 1977, and the first genome ever chemically synthesized, by Craig Venter’s team in 2003. It is also small enough that an AI trained on viral genomes has a realistic chance of writing new versions that actually work.
The AI systems involved are Evo 1 and Evo 2, genome language models developed by Brian Hie’s group at Stanford and the Arc Institute. Evo 2, with 40 billion parameters trained on roughly 9.3 trillion nucleotides, is the largest openly released genome model to date. For this project, the models were fine-tuned on 14,466 sequences from the Microviridae family, the group to which phi-X174 belongs, then prompted to generate complete genomes in a single left-to-right pass, writing the entire sequence the way a language model writes a sentence, without stitching together parts of existing genes.
The pipeline involved filtering thousands of candidate genomes, selecting roughly 300 for chemical synthesis based on predicted features, and testing each in the lab. The success rate, about 5%, is low in absolute terms but significant as a proof of concept: 16 of the synthesized genomes produced functional phages that could infect and kill E. coli. Several of the AI-generated phages outperformed the natural parent strain in growth competitions, and a cocktail of the 16 phages rapidly overcame E. coli strains that had evolved resistance to the natural virus, something the natural phages alone could not do. One generated phage, dubbed Evo-phi-36, incorporated a DNA packaging protein from an evolutionarily distant phage, a combination that does not exist in nature.
The biosecurity dimension is why the paper shipped with an unusually pointed companion essay. Tom Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security argue in Science that the ability to compose viral genomes with generative AI now exists while the governance to steer it does not. They note that AI-designed genomes might encode pathogens that cannot be contained by existing countermeasures, and that the researchers’ decision to exclude viruses infecting humans, plants, and animals from the training data can be partly circumvented by fine-tuning models on pathogen data. They call for legally required screening of synthetic DNA orders rather than reliance on voluntary safeguards.
The research team itself flagged the risks. The authors deliberately excluded human and animal viruses from training, urged others to consult safety and security professionals, and emphasized that the work was conducted under biosafety oversight. Independent experts were divided on the magnitude of the threat. Tom Ellis of Imperial College called the work impressive but noted that phi-X174 is the smallest and easiest genome to make, and argued the threat from full AI design of a dangerous genome is overblown compared with gain-of-function modifications of existing pathogens. Others pointed out that the roughly 5% success rate and the fact that natural evolution polished half of the functional phages suggest the models are still generating many nonfunctional sequences, a buffer that may not persist as the models improve.
A preprint of this work circulated in September 2025, and the current publication is the peer-reviewed version, with the accompanying governance analysis. The distinction matters: the models wrote the sequences, but humans chose the training data, designed the filters, selected which genomes to synthesize, and tested them. The innovation is that the writing itself, the generation of complete functional genomes, has moved into the machine.
The near-term promises are medical: phages that kill antibiotic-resistant bacteria, engineered viruses that deliver genes, biological tools built to order. The risks are the ones that have shadowed synthetic biology since its birth, now accelerated by models that make genome design cheap and accessible. The Perspective’s authors said the technology has arrived and the question is whether the rules will be written before the capability spreads.
Sources: King, S.H., Driscoll, C.L., Li, D.B. et al. Generative design of bacteriophages with genome language models. Science 393, eaec2657 (2026). DOI: 10.1126/science.aec2657. Inglesby, T.V. & Hanke, M.S. AI-designed viral genomes. Science 393, 563-564 (2026). DOI: 10.1126/science.aej8512. The Guardian, August 6, 2026.