AI engineers 16 novel virus genomes from scratch in first-of-its-kind study
The resulting viruses pose no threat to humans, though experts warn of unresolved questions regarding the safety of AI-designed biological entities.
Scientists at Stanford University and the Arc Institute have successfully trained artificial intelligence to design 16 novel virus genomes from scratch, marking the first time such a feat has been achieved without copying existing templates. The research, which utilised the bacteriophage Phi X-174 as a foundational basis, demonstrates the capacity of AI models to understand complex evolutionary patterns and biological constraints necessary for an organism to remain functional.
The AI systems employed in the study, identified as Evo 1 and Evo 2, were trained on millions of genomes across all domains of life. This extensive training allowed the algorithms to recognise how genes are typically organised and to identify the specific biological limits required for viability. By focusing on these underlying structures rather than replicating known viruses, the researchers were able to engineer new bacteriophages—viruses that infect bacteria and pose no threat to humans.
While the immediate results involve non-pathogenic entities, the implications of the technology extend beyond basic biology. The study suggests that this approach could potentially lead to significant medical breakthroughs. However, it also raises concerns regarding the possibility of the technology being used to develop biological weapons, a risk that remains speculative but significant in the broader context of AI safety.
Dr Moritz Hanke, a fellow at the Johns Hopkins Center for Health Security, highlighted the current lack of agreement on the risks associated with this technology. He told the New York Times that there is no consensus on the dangers of AI-designed viruses, underscoring the need for careful oversight as the field advances. The research builds on prior work involving AI engineering of bacteriophages but distinguishes itself by using de novo design methods rather than simple modification of existing genetic material.
The findings were reported in the MIT Technology Review newsletter The Download on 7 August 2026, summarising coverage from international outlets including the BBC, Wall Street Journal, Axios, and Wired. The study represents a significant step in the application of artificial intelligence to biological research, moving beyond analysis to active design of genetic structures.
