Researchers at Stanford University and the Arc Institute used an artificial intelligence model called Evo to create viable viruses for the first time, according to The New York Times. The model generated virus genomes that successfully replicated inside bacteria and infected them.
Unlike text-based systems, Evo trained on genetic sequences from millions of animals, plants, microorganisms, and viruses. In total, the model analyzed roughly nine trillion nucleotides. The New York Times noted that DNA functions like text, using molecular letters A, C, G, and T to store instructions for creating proteins and other molecules, and researchers wanted to test whether the neural network could independently learn the rules governing genetic sequences.
After learning gene patterns, scientists tested whether Evo could create entire genomes rather than individual genes. For their initial test, they chose bacteriophage Phi X-174, a virus that infects E. coli bacteria and has been studied for nearly a century. Stanford University doctoral student Samuel King, a co-author of the study, said the experiment seemed like the obvious next step.
Genome generation and lab testing
Following additional training on 11 genes from Phi X-174 and about 15,000 related viruses, Evo generated roughly 700,000 potential genomes. Researchers selected the most promising options, synthesized DNA for 285 model sequences, and inserted them into bacteria.
Most initial experiments failed, but clear spots appeared in one dish, signaling active viral replication. Further testing confirmed that 16 genomes created by Evo produced viable viruses, with some reproducing faster than natural Phi X-174, according to the study authors.
The publication noted that the created viruses pose no threat to humans. Developers did not train Evo on genetic data from human-infecting viruses, and dataset developers excluded related viruses capable of infecting animals, plants, and fungi.
Medical potential and safety concerns
Scientists see potential practical uses for the technology in medicine and biotechnology. Experts said medical treatments already use viruses as carriers to deliver genes to cells for genetic diseases, and AI could help create new tools if similar models work on other virus groups.
However, the experiment heightens concerns about the potential use of artificial intelligence to create dangerous pathogens. Moritz Hanke, an expert at the Johns Hopkins Center for Health Security, said technology development is outpacing the creation of safeguards.
The New York Times reported that while government rules address high-risk biological research, computer-generated genomes do not always fall under existing restrictions. Hanke stated that the main problem remains a lack of clear criteria to assess the dangers of entirely new viruses.
Artificial intelligence limitations
In additional artificial intelligence reporting cited by UNIAN, researcher Peter Denning argued in his book Turing's Error that modern systems will never surpass humans in complex thinking due to a lack of tacit knowledge. He defined tacit knowledge as common sense, practical skills, intuition, nonverbal communication, and cultural context built from experience.
Denning noted that neural networks process basic word meanings without accounting for connotations or context, allowing artificial intelligence to act quickly and powerfully while remaining unpredictable and dangerous.
