When a language model talks to humans, it predicts the next word. When a genome model talks to biology, it predicts the next nucleotide — and from there it can draft entire genomes that have never existed in nature. That is exactly what researchers affiliated with the Arc Institute and Stanford University demonstrated when they used Evo, a genomic foundation model, to design unprecedented bacteriophages. The best-known test case was ΦX174, a phage that infects E. coli: The system produced complete, fully functional sequences that were able to replicate in the lab. Sixteen of those designed viruses turned out to be viable and capable of infecting bacteria — a milestone that places generative artificial intelligence at the center of synthetic biology.
The idea behind these models is elegant. Just as ChatGPT learns statistical patterns from trillions of words, Evo and its successor Evo 2 were trained on trillions of DNA base pairs spanning every domain of life. Instead of text, they learn the grammar of evolution: how genes are organized, how proteins fold, and how whole genomes become viable. Published in Nature in 2026, Evo 2 extended this capability past 9 trillion bases, making it possible to model DNA, RNA and proteins at single-nucleotide resolution. It is the difference between memorizing isolated words and grasping the syntax of a full paragraph — except the paragraph is a chromosome. And just as chatbots can be fine-tuned for narrow tasks, these models can be steered to favor particular traits, such as the proteins on a virus's surface or the host it prefers to infect.
The ability to generate viral genomes is at once a promise and a dilemma. The so-called dual-use problem is not new in biology, but AI accelerates it: the very tool that can design a therapeutic phage to fight antibiotic-resistant bacteria could, in principle, be used to reconstruct or enhance pathogens of interest to hostile actors. That is why the biosecurity debate has gained fresh urgency. The authors themselves acknowledge the risk and argue for governance: screening sequences before DNA synthesis, ethical review of projects, and responsible access policies for the open model. In that context, international debates are beginning to treat biological foundation models as critical infrastructure, deserving oversight similar to that applied to high-containment laboratories.
Yet the constructive side is enormous. Designed phages open the door to phage therapy — a real alternative to antibiotics amid the crisis of bacterial resistance. Genome models also help predict how emerging viruses evolve, speed up vaccine design, and deepen our understanding of evolutionary history. Preparing properly for the next pandemic is not only about monitoring what exists, but anticipating what could arise.
Governance, however, is still chasing the technology. Today there is no clear international consensus on how to regulate the synthesis of AI-created genomes, and much of the rulemaking relies on voluntary standards from DNA synthesis companies. The open question is uncomfortable: once the technical barrier to creating a virus is no longer the bottleneck, will the thing that protects us be a human decision — or will we discover too late the price of not having made one?
Sources: Ars Technica, Engadget, Olhar Digital, Nature (Evo 2), Arc Institute (Evo), bioRxiv (Evo phages)
✓ Independent sources cross-checked and verified before publishing
