AI Generates 700,000 Novel Bacteriophage Sequences, 16 Prove Functional

A new scientific study has demonstrated that specialized artificial intelligence models trained on genomic data can design novel viruses capable of infecting bacteria and replicating inside them. The research, published in the journal Science, was reported by The New York Times and conducted by scientists from the Arc Institute in Palo Alto and Stanford University.

How Evo Models Learn the “Language” of DNA

The researchers employed AI systems named Evo 1 and Evo 2, which operate similarly to large language models such as ChatGPT and Claude—but instead of being trained on books or websites, they were trained on genetic sequences. Through exposure to trillions of nucleotides—the fundamental units of DNA—the models learned patterns and structural rules embedded in genetic code.

In the experiment, the team further trained the Evo models on approximately 15,000 viruses belonging to the same family as Phi X-174, a small virus that exclusively infects Escherichia coli. Scientists imposed strict experimental constraints: the models were explicitly instructed to exclude any viral sequences with potential to infect humans, other animals, plants, or fungi.

From 700,000 Candidates to 16 Functional Viruses

The Evo model generated roughly 700,000 candidate viral sequences. Researchers selected 285 of these as the most promising for laboratory testing. After synthesizing the DNA corresponding to those candidates and introducing it into bacterial cells, 16 of the Evo-designed constructs yielded functional, replicating viruses.

Analysis showed these AI-generated viruses were at least as viable as the naturally occurring Phi X-174 virus. Some demonstrated faster replication rates than Phi X-174 under experimental conditions.

Potential Applications and Escalating Biosafety Concerns

This capability holds significant promise for science and medicine: AI-designed viruses could advance gene therapy development, deepen understanding of genome function, and support creation of more efficient tools against harmful bacteria. Yet the same technical capacity raises serious biosafety concerns—particularly given the rapid pace of AI model advancement.

A system capable of designing new genomes could, in malicious hands, become a far more dangerous instrument than AI models merely describing biological threats. Future iterations of such models may gain enhanced ability to engineer biological entities with hazardous properties.

Regulatory Lag and the Dual-Use Dilemma

The experiment underscores a growing challenge in AI governance: the technology itself is neutral, but its application can yield either transformative benefits or unprecedented risks. The study’s authors implemented clear safeguards—including restricting targets exclusively to bacteria and excluding all sequences with cross-species infection potential.

However, broader questions remain unresolved: what happens as such capabilities migrate to more advanced models, and whether existing biosafety oversight frameworks can keep pace with the speed of AI development? While AI-driven virus design may mark a critical step toward novel therapeutics and biological tools, it simultaneously demands stringent controls to prevent the shift from scientific research tool to biological threat generation platform.