An artificial intelligence model built using technology from OpenAI and trained by researchers at Stanford University designed entirely new viral genomes that were functional when synthesized in a laboratory, according to a study published Thursday in the journal Science.
AI-created genomes proved infectious in bacteria
The generative model, named Evo, was trained on massive collections of DNA sequence data — described by the authors as comprising millions of genomes — to learn the evolutionary constraints that shape natural DNA. Researchers then used Evo to produce novel sequences that had not been observed in nature. After chemical synthesis and laboratory testing, 16 of the AI-designed bacteriophages successfully infected Escherichia coli, including some that appeared to overcome the bacteria’s native resistance mechanisms.
| Item | Detail |
|---|---|
| Model | Evo (OpenAI-based generative model) |
| Training data | Millions of genomes (DNA data) |
| Laboratory result | 16 synthesized phages infected E. coli |
| Human pathogen data | Excluded from training sets |
Scientific promise and security concerns
The authors of the paper explicitly excluded datasets containing human pathogens from Evo’s training material, a safeguard intended to limit the model’s ability to design agents that could infect people. Nevertheless, the work highlights both a potential tool for addressing antibiotic resistance and a renewed set of risks for biosecurity experts.
Proponents say the approach could accelerate the discovery of bacteriophages that target antibiotic-resistant bacteria, enabling tailored therapies rather than relying solely on viruses found in nature. That potential therapeutic benefit sits alongside warnings from biosecurity leaders that the research demonstrates how quickly AI can be harnessed to produce novel biological agents.
In an accompanying commentary in Science, biosecurity experts at the Johns Hopkins Center for Health Security — Dr. Thomas Inglesby and Dr. Moritz Hanke — said the study shows AI is now capable of generating biological constructs that raise real risks and called for stronger legal measures and oversight for future work.
- Research demonstrates generative AI can learn evolutionary patterns from large DNA datasets and create functional genomes.
- Laboratory synthesis and testing verified infectivity of multiple AI-designed bacteriophages.
- Authors excluded human pathogen sequences from training, but experts warn broader policy gaps remain.
What this means for regulation and research
The paper arrives amid growing debate over the governance of advanced AI tools applied to biology. Policymakers, public health officials and the scientific community face difficult trade-offs: enabling scientific advances that could address urgent problems such as antimicrobial resistance while preventing misuse that could threaten public health.
Experts who study dual-use research — studies that can be used for both beneficial and harmful ends — have urged clearer frameworks for who can access high-risk biological design tools, how models are evaluated before publication, and what safety measures must be in place before lab work proceeds. The Stanford study is likely to intensify those conversations by showing that a model trained on sequence data can output previously unseen, functional viral genomes.
The authors and journal underscored that the study did not aim to create human pathogens and that key safeguards were implemented. Still, its results underscore how rapidly capabilities at the intersection of machine learning and synthetic biology are advancing, and how existing norms and regulations may need to be reassessed to keep pace.
As regulatory and biosecurity discussions continue, scientific institutions and funding agencies will face pressure to balance openness and innovation with risk mitigation. The paper serves as a concrete example that may drive new policy proposals at the federal level and within the broader international public health community.
