![Cover image for [Scientists Have Created Viruses Designed by AI] But What That Really Means?](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqgstfcfbn5lua7ywpdgu.png)
The viruses cannot infect people. In fact, they may eventually help save lives. But the experiment marks a turning point: artificial intelligence can now help design complete biological systems that reproduce and evolve.
The phrase “viruses designed by artificial intelligence” sounds like the opening scene of a disaster film.
The reality is far less apocalyptic—but no less important. We know how things starting...
Scientists from Stanford University, the Arc Institute and the Broad Institute have used generative AI to design complete viral genomes. They then built hundreds of those proposed genomes in a laboratory. Sixteen worked: they became functioning viruses able to infect bacteria, reproduce and continue evolving.
These were not human viruses. They were bacteriophages, or phages—viruses whose natural hosts are bacteria. The researchers tested them against harmless laboratory strains of Escherichia coli, not people or animals.
No new human disease was created. No AI spontaneously brought an organism to life. And the experiment still required skilled researchers, DNA synthesis, careful screening and extensive laboratory work.
Even with those qualifications, something remarkable happened.
For years, scientists have used AI to analyze DNA, predict protein structures and suggest individual molecules. In this experiment, AI helped design an entire functioning genome: a coordinated biological system containing multiple genes and regulatory elements that must work together for the virus to survive.
Biology has entered a new phase. We are beginning to move from reading and editing the code of life to generating it.
Why would anyone ask AI to design a virus?
The experiment had two purposes.
The first was scientific.
Researchers wanted to know whether a genome language model could learn enough about biology to design something more complicated than an individual protein.
These models work in a way loosely comparable to the language models behind chatbots. A language model learns statistical relationships between words and uses them to generate new text. A genome model instead studies sequences written in DNA’s four-letter alphabet: A, C, G and T.
But generating DNA that looks convincing is not the same as generating DNA that works.
A complete genome must coordinate many activities. It has to produce the correct proteins, copy itself, recognize a suitable host and package its genetic material. Regulatory signals must operate at the right time. Proteins must fit together. A single badly placed mutation can make the entire genome useless.
The researchers chose a phage called ΦX174 as their starting point. It is one of the best-studied viruses in science and has a tiny genome containing 5,386 DNA letters and 11 genes.
Small does not mean simple. Several of its genes overlap, so the same section of DNA can help encode more than one protein. A change that benefits one gene may damage another. That made ΦX174 manageable enough to synthesize but difficult enough to provide a serious test.
The virus also has a special place in the history of genetics. In 1977, ΦX174 became the first organism to have its complete DNA genome sequenced. In 2003, scientists used it to demonstrate that a genome could be assembled from chemically synthesized DNA.
Now it has become the template for another milestone: the first complete functional genomes designed using generative AI. Arc Institute
The second purpose was medical.
Phages infect and frequently kill bacteria. That makes them possible weapons against bacterial infections, particularly when antibiotics no longer work.
Antimicrobial resistance is already a global health crisis. The World Health Organization estimates that drug-resistant bacterial infections directly caused about 1.14 million deaths in 2021 and were associated with 4.71 million deaths. Current projections suggest antimicrobial resistance could directly cause 39 million deaths between 2025 and 2050. WHO
Phage therapy is not a new idea. It was explored before antibiotics became widely available and is still used experimentally or in limited clinical settings today.
Its central problem is evolution. A phage may initially kill a bacterial strain, but the bacteria can mutate and become resistant. Researchers must then search for another phage or modify the one they already have.
AI could make that search faster. Instead of looking through oceans, soil or sewage for a suitable natural virus, scientists might one day generate a large library of phages tailored to attack a particular bacterium.
How the viruses were made
The researchers used genome language models called Evo 1 and Evo 2.
The original Evo models had learned from more than two million phage genomes. The team then gave the models additional training using 14,466 sequences from Microviridae, the family of small viruses that includes ΦX174.
What followed was not a one-click process.
The AI generated thousands of possible genomes. Researchers developed software to examine their genes, organization and similarity to natural phages. They rejected sequences that appeared incomplete or biologically implausible and selected candidates with features likely to preserve the desired host range.
They eventually chose 285 designs to build and test physically.
To turn a digital sequence into a virus, the scientists first had to synthesize its DNA. That DNA was assembled and placed inside bacterial cells, which provided the machinery needed to produce viral particles. The researchers then watched to see whether the bacteria stopped growing—evidence that a functional phage had emerged.
Sixteen candidates worked.
That is only about 5.6% of the 285 physically tested designs, and a far smaller share of the thousands initially generated. Most of the AI’s proposals therefore failed.
This is an important reality check. The experiment did not demonstrate that AI can effortlessly produce a working virus on demand. Independent scientists described it as an impressive but inefficient proof of concept. One suggested that the unsuccessful genomes could be considered biological versions of an AI “hallucination”: convincing-looking outputs that fail when tested against reality. Science Media Centre
But the 16 successes were real.
Each was genetically different from known natural phages. They contained between 67 and 392 mutations compared with their closest identified relatives. Thirteen carried mutations that had not been observed in known natural sequences.
One, called Evo-Φ2147, shared only 93% of its DNA with its closest known relative. Under some classification systems, that could make it a new viral species.
Another phage, Evo-Φ36, contained a DNA-packaging protein normally found in a distantly related virus. Previous attempts to transplant that protein through conventional engineering had failed. The AI-generated genome appears to have included other changes that helped the unfamiliar component work inside its new system.
That result hints at the real power of generative biology. Human engineers often change one component at a time because predicting many interacting mutations is extremely difficult. AI can explore combinations across an entire genome and occasionally find arrangements that a human researcher might never consider.
It does not necessarily understand why those arrangements work. But it can help find them.
The phages took on resistant bacteria
The team next allowed E. coli to develop resistance to the natural ΦX174 phage.
The bacteria changed parts of their surface, making it more difficult for the virus to attach and infect them. ΦX174 could no longer control the resistant strains.
The researchers then exposed those bacteria to cocktails containing the AI-generated phages. The new viral populations overcame resistance after further rounds of infection and evolution.
Some successful phages became genetic mosaics, combining material from two or three AI-generated designs through natural recombination. AI had created a diverse set of starting points; evolution then mixed and refined them.
This combination may be more important than any individual virus. Rather than trying to design one perfect treatment, researchers could generate a population with many different routes for attacking a bacterium. If the bacterium escapes one phage, another may still work.
That could eventually turn phage therapy from a slow hunt for suitable viruses into a more systematic design process. Original Science study
Eventually is the key word.
The new phages have not been shown to treat an infection in a human being. They have not undergone clinical trials. Their safety, stability and effects on the human microbiome remain unknown.
A phage that works against a laboratory strain of E. coli may fail against bacteria taken from actual patients. The immune system may remove it too quickly. Bacteria may develop new forms of resistance. Manufacturing a consistent medicine from viruses that can evolve presents another challenge.
The study is a beginning, not a therapy.
What the headlines leave out
It is technically correct to say that scientists created viruses designed by AI. But without context, that description encourages several false impressions.
First, these viruses cannot infect people.
The researchers used harmless laboratory bacteria, and the phages retained a narrow host range in the strains they tested. Bacteriophages are not miniature versions of human viruses waiting to switch targets. The biology required to infect a bacterial cell is profoundly different from the biology required to infect a human one.
Second, the AI did not conduct the experiment by itself.
Human researchers chose the virus, assembled the training data, adjusted the model, defined the design constraints, filtered thousands of candidates, synthesized the DNA and tested every design. Most candidates failed.
Third, the model had been deliberately trained without viruses that infect humans, animals or plants. The Evo developers report that efforts to prompt the model for human viral sequences produced essentially random outputs.
That safeguard matters. But it does not mean AI-designed human pathogens will always be impossible.
Another organization could train a model on different data. Future systems could be more capable. Restrictions could be removed. Biological models might also become connected to automated laboratories, reducing the amount of manual work required.
The experiment is safe in its immediate form. The underlying direction of travel is harder to dismiss.
Why biosecurity experts are concerned
The study was accompanied by a Science article from Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security.
Their argument was not that these 16 phages present a danger. It was that the ability to generate complete functional viral genomes now exists, while the systems for governing that ability remain incomplete. Johns Hopkins
A future model trained on more sensitive material might propose viruses of humans, animals or crops. Some designs could be different enough from known pathogens to complicate existing diagnostic tests, vaccines or treatments.
There are several possible forms of risk.
An accidental risk could arise if researchers fail to anticipate a generated virus’s properties or work without adequate containment.
An institutional risk could arise if laboratories pursue increasingly ambitious experiments without independent safety review.
And there is the risk of deliberate misuse. AI might eventually help a state or a technically capable group modify an existing pathogen more quickly, explore combinations of mutations or evade biological defenses.
That does not mean AI is currently the easiest way to create a dangerous virus. Brian Hie, the study’s senior author, has argued that modifying an existing pathogen through conventional genetic engineering would be easier and more attractive to a malicious actor than designing one from scratch.
Even an AI-generated sequence must still cross the physical world. Someone needs the necessary expertise, suitable laboratory facilities, synthesized DNA, host cells and a way to test failed designs. Those remain significant barriers.
They may not remain equally significant forever.
Safety cannot exist only inside the AI
One of the clearest lessons from the study is that no single safeguard will be enough.
Removing dangerous sequences from training data can reduce risk, but another model may be built without that restriction. Controlling access to powerful models can help, but models can leak or be recreated. Laboratory rules matter, but they apply only to institutions that follow them.
The most credible approach is a series of overlapping barriers.
Research involving complete genome design should undergo independent biological-safety and security review. Models should be tested to determine whether they can assist with dangerous biological tasks. High-risk training data and capabilities may require access controls.
Laboratories still need physical containment, careful disposal procedures and transparent incident reporting.
DNA synthesis is another critical checkpoint. A generated genome remains computer data until someone manufactures the physical genetic material. Companies that sell synthetic DNA can screen orders and customers for warning signs.
That screening must become more sophisticated. Traditional systems often look for close matches to known pathogens. An AI-generated sequence could be biologically dangerous while looking substantially different from anything in a database. Future screening may need to identify hazardous functions, not simply familiar sequences.
In the United States, federally funded researchers are expected to purchase synthetic genetic material from providers that follow government screening standards. A 2025 executive order also called for broader, verifiable screening and a strategy addressing research outside federal funding. International coverage, enforcement and technical standards remain uneven. White House
Viruses, of course, do not respect national rules or borders.
What this could mean for the world
If the technology develops safely, its most immediate benefit could be a new generation of treatments for bacterial infections.
Scientists may be able to design phages against drug-resistant Pseudomonas aeruginosa, which can cause dangerous infections in hospitals and in people with chronic lung disease. Similar tools might target bacteria responsible for crop diseases, food contamination or harmful disruptions of the microbiome.
AI-generated genomes could also become research instruments. By comparing designs that work with those that fail, scientists may discover previously hidden rules about gene interaction, viral structure and evolution.
The broader change may be cultural. Biology has traditionally depended heavily on discovery: finding a useful organism, gene or molecule somewhere in nature and adapting it for human purposes.
Generative biology introduces another possibility. Scientists can begin with a desired function, create thousands of possible biological designs on a computer, synthesize the most promising ones and feed experimental results back into the model.
This could accelerate research dramatically. It could also concentrate power in institutions that possess advanced models, automated laboratories and large-scale DNA-synthesis capacity.
That raises questions about inequality. The countries most affected by antibiotic resistance may not be the first to receive personalized phage treatments. Yet they would share the consequences of any accident or misuse. Decisions about access, research standards and emergency preparedness therefore cannot be left entirely to a small group of companies or wealthy nations.
A warning—but not the warning people imagine
The lesson of this experiment is not that AI has created a pandemic.
It has not.
The new phages infect bacteria, were developed under controlled conditions and required substantial human expertise. Out of thousands of digital designs, only 16 became verified viruses. There is still an enormous distance between this study and the automated creation of a dangerous human pathogen.
The real warning is subtler.
A boundary has moved. Generative AI can now contribute to designing complete genomes that become physical, reproducing biological systems. The models will improve. DNA synthesis will become cheaper. Laboratories will become more automated. The knowledge required to connect those pieces will spread.
That same progress could give medicine powerful tools against antibiotic resistance and other diseases. It could help scientists explore forms of biology that evolution has never produced. It could also make biological mistakes or misuse easier to scale.
Those futures are not predetermined by the technology. They will be shaped by the choices surrounding it: what data models are trained on, who can access them, which experiments receive approval, how DNA orders are screened and how quickly the world invests in surveillance and medical defenses.
Humanity first learned to read DNA. Then it learned to rewrite it.
Now it is beginning to imagine genomes—and make some of them real.
Top comments (0)