AI is learning to design viruses. What could go wrong?
A Stanford and Arc Institute study shows that AI can generate functional bacteriophages, marking a new step in AI-driven biological research.
Artificial intelligence has taken another step into biology. Researchers from Stanford University and the Arc Institute have used AI to design complete bacteriophage genomes, with some of the resulting viruses successfully produced and tested in a laboratory.
The work shows that AI can move beyond analysing biological data and help create functional viral designs.
This research could eventually support new approaches to antibiotic-resistant infections. It also raises questions about how biological AI should be developed and governed as its capabilities grow.
AI-generated viruses enter the lab
The study, titled “Generative design of novel bacteriophages with genome language models”, was led by researchers including computational biologist Brian Hie. The team used genome language models, including Evo 1 and Evo 2, to generate new bacteriophage genomes.
Bacteriophages are viruses that infect bacteria rather than humans. In this study, the researchers focused on ΦX174, a bacteriophage that infects E. coli.
The team selected 302 AI-generated designs for synthesis and laboratory testing. Sixteen of those designs produced viable bacteriophages, meaning the AI-generated genomes could be turned into functioning viruses.
Some of the resulting phages also showed stronger performance than the natural ΦX174 in laboratory experiments. Researchers further tested combinations of the AI-designed phages against E. coli strains that had developed resistance.
That distinction is important. The study did not involve designing viruses to infect humans. It demonstrated that AI can help generate functional viruses that target bacteria.
Why the research matters
The biggest potential application is phage therapy, which uses bacteriophages to target harmful bacteria.
Antimicrobial resistance is making some infections increasingly difficult to treat with conventional antibiotics. Researchers have therefore been exploring phages as another way to attack bacteria.
AI could make this research faster by analysing genetic patterns and generating large numbers of potential candidates before scientists move to laboratory testing. Instead of relying entirely on natural phages found in the environment, researchers could eventually use AI to explore designs tailored to particular bacterial targets.
Where the risks begin
The same capability also creates a dual-use concern. Biological design tools developed for medical research could potentially be misused if similar techniques are applied to more harmful organisms. That does not mean the Stanford and Arc Institute study created a dangerous human pathogen. It did not.
The work focused on bacteriophages that infect bacteria. However, the experiment demonstrates an important capability: AI can generate biological designs that become functional after laboratory synthesis and testing.
As AI models become more capable, researchers will need to consider how such systems are accessed, what biological data they can use and how potentially risky designs are screened.
Decode Age sells gut microbiome tests and supplements built around the biology of ageing
The safeguards researchers need
DNA synthesis screening is one important layer. Companies that manufacture genetic sequences can check orders against databases of concerning sequences before producing them. Other safeguards could include controlled access to advanced biological AI systems, stronger risk assessments and independent testing of models used for biological design.
Clear documentation will also matter. Researchers should be able to establish how an AI-generated sequence was produced, what testing was performed and what safety controls were used.
What comes next
The research does not mean AI can freely create any virus scientists ask for. It does, however, demonstrate a meaningful shift in biological research. AI can now help generate complete viral genomes that can be synthesised and tested in the real world.
For medicine, that could eventually help researchers develop new ways to tackle resistant bacteria. For the wider AI industry, it is a reminder that biological capabilities need to advance alongside strong safety controls.
The technology is promising, but its next phase will depend on whether scientific progress and responsible oversight can move at the same speed.