The alarm over artificial intelligence building killer pathogens is louder than ever, fueled by headlines about genome-language models designing functional viruses. Yet infectious-disease experts and virologists warn that the true biological threats facing humanity still stem from natural evolution rather than machine learning algorithms.
Why AI-Made Pathogens Dominate the Headlines
Last month, a high-profile paper described a team of scientists teaching an AI model to help produce brand-new, deadly viruses. This genome-language model, trained on DNA sequences, designed multiple new genomes that researchers subsequently synthesized into functional viruses capable of killing bacteria in a dish. The study triggered immediate fear that the technology could manufacture pathogens capable of sparking a devastating human pandemic, echoing warnings raised by researchers and tech leaders like Bill Gates.
These anxieties intensified further when Anthropic released a report detailing several instances in which the company banned accounts using its LLM, Claude, for what could have been reputable research or attempts to develop biological weapons. This relentless pace of development has led many to fear that AI might soon make independent breakthroughs, potentially achieving recursive self-improvement where models design newer and smarter models.
Kevin Esvelt, a biologist at MIT and a co-founder of the biosecurity nonprofit Secure Bio, notes that while the probability of an AI-driven pandemic remains low, the potential consequences demand aggressive risk reduction. According to Esvelt, it is critical to drive that probability down as low as possible.
The Reality of Lab Bench Science Versus Silicon
Despite public anxiety, many infectious-disease experts and virologists argue that the threat of an AI-induced pandemic sits low on the list of actual perils. Viruses shaped by natural evolution still present immediate and far more pressing dangers. While bad actors could theoretically use AI to speed up infectious destruction, current technical hurdles remain immense.
The bacterium-killing virus created by researchers at Stanford University and the Arc Institute relied on an AI model called Evo, fed exclusively with simple bacterial-infecting viral genomes. Genomes targeting human-infecting viruses are vastly more complex. Brandon Ogbunu, an evolutionary biologist and infectious-diseases modeler at Yale University, points out that current AI models simply put a supreme battery behind methods researchers were already utilizing.
Furthermore, human understanding of pathogenic viruses remains severely limited. Virologist Seema Lakdawala of Emory University explains that current training data is essentially junk because generations of virologists have yet to answer fundamental questions about what makes certain viruses transmissible or why specific populations are more susceptible. Brian Hie, the lead researcher on the Evo team, remains unconvinced that AI would even be the method of choice for malicious actors, simply because the underlying knowledge base is lacking.
The Friction Between Biosafety and Public Health Research
While models powerful enough to suggest new viral genomes require strict review, excessive restrictions are already crippling legitimate scientific inquiry. Anthropic’s recent report captured routine scientific queries flagging safety filters, encompassing gain-of-function work on the chikungunya virus and studies on how flu viruses adapt to mammals.
Researchers like Lakdawala and Virginia Tech viral-transmission researcher Linsey Marr report that standard tools like Claude and ChatGPT frequently flag or halt daily academic conversations regarding basic genetic quirks or viral inactivation. To counter this, scientists are attempting to develop specialized frameworks to help AI models separate legitimate pursuits from bioterrorism.
Ultimately, AI remains a tool bound by human prompts and limited data. Gigi Gronvall, a health-security expert at the Johns Hopkins Bloomberg School of Public Health, emphasizes that AI cannot build or test physical pathogens. Successfully engineering a dangerous agent requires complex lab skills, specialized equipment, and animal models—hurdles far greater than a simple software prompt.
The Superior Efficiency of Natural Evolution
Nature remains the ultimate bioengineer. Viruses constantly mutate and test new iterations without human preconceptions about dangerous disease agents. Ogbunu observes that natural evolutionary processes operating across ecosystems are inherently more productive at generating novel variants than any current digital model.
Moving forward, the scientific community must balance necessary oversight with the urgent need to harness machine learning for vaccine development, viral surveillance, and epidemiological modeling. How should the scientific community draw the line between restricting dangerous biological queries and protecting the open research required to stop the next natural outbreak?
