AI Designed Working Viruses From Scratch, Scientists Confirm

Researchers at Stanford University and the Arc Institute have used an AI model to generate functional viral genomes, some of which outcompeted their natural counterparts once tested in living bacteria. The work, published in Science in August 2026, marks one of the clearest demonstrations yet that genomic language models can produce viable biological entities rather than just predict or classify them.

The team relied on Evo 1 and Evo 2, AI systems built the same way large language models like ChatGPT are, except trained on genetic sequences instead of text. Evo was first exposed to roughly nine trillion nucleotides spanning animals, plants, microbes, and viruses, then fine-tuned on the eleven genes of the bacteriophage Phi X-174 and about 15,000 of its closest relatives.

That narrow focus was deliberate. Phi X-174 and its relatives infect only E. coli bacteria, and the researchers excluded any training data on viruses capable of infecting humans, animals, plants, or fungi, a choice framed explicitly as a biosafety safeguard.

From 700,000 designs to 16 working viruses

Evo generated around 700,000 candidate genome designs. Researchers narrowed that pool down for physical testing, though the reported figures diverge slightly: one account puts the number of synthesized candidates at 285, while another cites approximately 302 genomes chemically manufactured as DNA and inserted into bacteria. Both figures come from the same body of reporting on the study, and the discrepancy likely reflects different counting stages in the pipeline rather than a factual dispute.

Of those candidates, 16 produced viruses that successfully replicated inside bacterial cells. Notably, some replicated faster than the naturally occurring Phi X-174 virus, meaning the AI didn't just replicate biology, it improved on it in a measurable, competitive sense.

A regulatory gap the researchers acknowledged

The timing intersects with a fresh policy question in the United States. Late July saw the National Institutes of Health issue new rules on high-risk life sciences research, explicitly banning experiments designed to make pathogens more dangerous. But that policy stops short of covering purely computational design work, such as generating viral DNA sequences on a computer, unless a flagged "entity of concern" is involved.

That leaves AI-driven genome design sitting in a gray zone: the physical synthesis and lab testing steps may fall under existing biosafety oversight, but the generative modeling itself currently does not.

The researchers restricted their model to bacteria-only viruses specifically to keep the demonstration far from anything with human or agricultural health implications.

The result is a proof of concept rather than a tool for engineering dangerous pathogens, since the model was intentionally starved of any training data related to human, animal, or plant pathogens. But it establishes that generative AI can now cross from sequence prediction into functional biological creation, a capability that regulators, biosecurity researchers, and AI labs will need to grapple with as these models scale beyond bacteriophages.