Artificial intelligence has crossed another line.
Researchers at Stanford University have used generative AI to design complete viral genomes, with 16 of the resulting designs successfully produced as functioning bacteriophages, viruses that infect bacteria rather than humans.
The immediate goal is promising: develop new weapons against dangerous bacteria, including infections that have become resistant to antibiotics.
But the achievement raises a much larger question.
If AI can learn to design useful viruses today, what will increasingly powerful systems be capable of designing tomorrow?
That concern is no longer confined to science fiction. Biosecurity specialists are already warning that the convergence of AI and biotechnology could eventually make dangerous biological engineering easier.
From Predicting Words to Predicting Life
Most people know generative AI through systems that produce text, images, video or computer code.
Biological AI operates in a different language: DNA.
The Stanford researchers used models known as Evo 1 and Evo 2, which are trained on biological sequences and can predict and generate genetic code.
Researchers then tasked the technology with designing bacteriophages targeting E. coli. Hundreds of potential genomes were generated, and 16 designs ultimately produced viable viruses, according to the research.
That represents an important scientific milestone: generative AI moving beyond designing individual proteins or pieces of biological machinery toward designing an entire functional genome.
“This is a next step in the complexity that’s designable by generative AI,” Stanford assistant professor Brian Hie told the BBC, describing it as new territory for his team.
And unlike a computer-generated picture or paragraph, these designs can become physical biological entities capable of replicating.
That is what makes the breakthrough extraordinary. It is also what makes it potentially consequential.
There Is a Very Good Reason Scientists Want This Technology
Bacteriophages are not inherently sinister. They infect bacteria, and scientists have studied their potential medical uses for decades.
One particularly important possibility is phage therapy.
As bacteria evolve resistance to antibiotics, infections that modern medicine once treated relatively easily can become increasingly difficult to fight. Scientists hope carefully selected or engineered bacteriophages could provide another way to attack those bacteria.
AI could potentially accelerate that process dramatically.
Instead of relying exclusively on naturally occurring phages, scientists could eventually use computational systems to help design biological tools tailored to specific bacterial targets.
That could lead to important medical advances.
But powerful technologies rarely come with only one possible application.
The Same Capability Creates a Biosecurity Problem
The Johns Hopkins Center for Health Security has been studying precisely this intersection of artificial intelligence and biotechnology.
Its experts say biological AI could revolutionize medicine, vaccine development, diagnostics and outbreak detection. But they also warn that sufficiently capable systems could simplify the creation or modification of dangerous biological agents.
That is the uncomfortable reality behind the Stanford breakthrough.
The researchers designed viruses that attack bacteria, not humans. There is a vast difference between producing a bacteriophage under controlled laboratory conditions and producing a novel human pathogen.
But a capability has now been demonstrated.
AI can help generate a complete viral genome that can subsequently function in the physical world.
The argument over whether AI will ever participate in designing complete viruses has therefore changed. The more important question is what safeguards should surround increasingly powerful biological-design systems.
Researchers Say They Built Safeguards Into the Experiment
Hie emphasized that his team attempted to minimize the danger.
According to his account, viruses capable of infecting complex organisms were excluded from the relevant training data, and the experiments were limited to bacteriophages targeting bacteria. The work was also performed under laboratory controls.
Those distinctions matter.
This was not an experiment in designing a new human disease, and describing it that way would badly misrepresent the research.
But biosecurity experts aren’t necessarily worried about these 16 bacteriophages.
They’re worried about where the technology goes next.
Johns Hopkins researchers have previously recommended heightened scrutiny for biological AI models with concerning capabilities or those trained on particularly sensitive biological information. They have also described nucleic-acid synthesis screening as an important final safeguard against biological misuse.
In other words, one defense is preventing an AI system from producing dangerous designs. Another is preventing dangerous designs from being physically synthesized.
As AI improves, both layers could become increasingly important.
This Isn’t a Hypothetical Concern Anymore
The Johns Hopkins Center specifically identifies two high-consequence possibilities that deserve governance: AI substantially simplifying the recreation of dangerous viruses and AI making it easier to create novel pathogen variants or biological constructs capable of producing catastrophic outbreaks.
That doesn’t mean today’s Stanford system can do those things.
It means experts can see the trajectory.
And biological research already contains examples where scientists themselves have concluded that technical capability should not automatically translate into experimentation.
In 2024, a group of 38 experts, including Johns Hopkins Center director Thomas Inglesby and two Nobel laureates, called for researchers to refrain from attempting to create hypothetical “mirror bacteria” because of the potentially extraordinary consequences if such organisms escaped control.
The principle is straightforward: “Can we?” and “Should we?” are two different scientific questions.
AI is making the first question easier to answer.
Society may have to get much better at answering the second.
AI Is Leaving the Screen
That may ultimately be the biggest story here.
For years, debates over artificial intelligence centered on jobs, misinformation, surveillance, deepfakes and political bias.
Those remain important.
But biological AI introduces a fundamentally different category of risk because its outputs don’t necessarily remain digital.
An AI-generated article exists on a screen.
An AI-generated image consists of pixels.
An AI-generated biological sequence can potentially become something physical.
Something that reproduces.
That does not make biological AI inherently dangerous. The same capabilities could help scientists develop medicines, defeat antibiotic-resistant bacteria and respond faster to emerging diseases.
But the Stanford experiment demonstrates why biosecurity needs to develop alongside the technology rather than years behind it.
The breakthrough isn’t that AI suddenly created a pandemic virus. It didn’t.
The breakthrough is that researchers have demonstrated generative AI can help design complete, functional viral genomes.
That’s an extraordinary scientific achievement.
And it’s exactly the kind of achievement that deserves serious attention before the technology becomes far more powerful.