Last Updated: 8 August 2026
Stanford researchers built complete bacteriophage genomes using two AI models: Evo 1 and Evo 2. The team generated thousands of candidate sequences. Nearly 300 got synthesized and tested in the lab. Sixteen actually worked against E. coli. Stanford recently announced the results, though the underlying research had been done for a while.
Earlier AI tools could only tweak a single gene or make a small edit here and there. This one went further and built an entire virus genome from start to finish, on its own. So did the AI actually "create" a virus, or is that claim bigger than what really happened? It's a question worth digging into.
The AI models produced complete genetic blueprints. Researchers then took some of those blueprints and turned them into real, physical viruses. In the end, testing produced 16 that actually worked. This is described as the first time AI has designed a working virus genome from scratch, based on a real virus called ΦX174.
Stanford says the AI came up with thousands of possible genome designs. Researchers picked a much smaller group from that pool to actually build and test in the lab. Close to 300 of them were built and tested in the end.
It helps to separate two things here: what the AI made, and what became a real virus. The AI's output was just data, a DNA sequence on a screen, not an actual virus. Turning that data into something biological took a separate process, done entirely by people, which we'll walk through next.
Evo generated complete viral genome sequences rather than suggesting individual mutations one at a time, but it worked from information tied to ΦX174 and remained one part of a larger, human-supervised design and testing pipeline. That distinction matters for understanding what "AI-designed" really means here.
Stanford says Evo 2 could start from a small snippet of ΦX174 DNA and generate novel candidate genomes from it. Bryan Hie, the researcher behind the work, said the genomes were generated in a single left-to-right pass rather than pieced together or completed manually afterward. The research abstract confirms that Evo 1 and Evo 2 were used to generate whole-genome sequences with desired biological characteristics.
Generating a genome sequence is not the same as producing and validating a functional biological virus. ΦX174 served as the design template throughout, so this was not generation "from nothing." The model was working with biological context, not inventing a virus in a vacuum.
Getting from an AI-generated sequence to a real virus took a multi-stage process: AI generation, computational evaluation, candidate selection, chemical DNA synthesis, and laboratory testing. Each stage depended on human decisions.
Evo produced thousands of genome candidates. Samuel King developed a computational framework to evaluate those candidates against design criteria based on ΦX174 and related phages, since testing every single candidate in a lab would have been impractical. Stanford notes that high DNA-synthesis costs were part of the reason this computational filtering mattered so much.
Once the framework narrowed the field, researchers selected promising candidates rather than synthesizing everything Evo produced. Those selected sequences were chemically synthesized into DNA, and the resulting genomes were tested experimentally to see which ones produced viable, effective phages. In short: AI generation, computational filtering, human selection, DNA synthesis, and laboratory validation, in that order.
Researchers picked ΦX174 for a simple reason: it's small. That made it easier to work with and easier to test. But it still gave them a real answer to the question they cared about, could AI actually design a virus that works and kills bacteria? ΦX174 attacks E. coli, and its genome runs under 6,000 letters.
For comparison, the human genome has about 3 billion letters. So ΦX174 is nothing next to that. But don't let "small" fool you. Even 5,400 letters is more than a person can just read and make sense of. That's the gap AI was brought in to close. Here's the thing: even a "simple" virus isn't simple to design by hand. There's too much to track, too many possibilities. AI can look at that same information and build something workable in a fraction of the time it would take a person.
Every virus the AI generated in this study was built off ΦX174 as a template. Worth remembering though: pulling this off on something tiny doesn't prove it'll work on something bigger or messier. That's still an open question.
All 16 phages infected E. coli, just like they were designed to. A few even beat the natural virus they were based on. And when researchers mixed several of them together, that blend took down bacteria that had already figured out how to resist the original virus on its own.
Some of these lab-made phages grew faster than the natural one. They also broke open bacterial cells quicker. So instead of picking just one winner, researchers grabbed a handful of the strongest phages and combined them. That mix didn't just work once, it beat three different resistant strains of E. coli in testing.
Stanford is calling this a proof of concept, and that's fair. It shows the idea holds up, not that it's ready for anything real-world yet. The bacteria used here had already beaten the natural virus before this test even started. The AI-built cocktail still got through, which is what the team was hoping to see.
The logic behind mixing phages is simple: bacteria have a harder time resisting several attackers at once than just one. But keep this in perspective. This happened in a lab, not a hospital. It's a good sign for phage therapy down the road, but there's no treatment here yet, just early evidence.
Evo came up with the genome designs. That part was all AI. But everything after that was still up to people. Researchers picked which designs were worth trying. They built the actual DNA. They ran the lab tests. They looked at the results and figured out what they meant. Evo never touched a test tube.
Evo couldn't build DNA on its own. It couldn't turn its own designs into real viruses either. A tool built by the research team decided which designs looked promising enough to try. From there, people made the final choice on what to actually build. Then lab work, done by hand, showed which ones worked and which ones didn't.
That doesn't mean researchers just did the work themselves and gave Evo the credit. Coming up with full, working genome designs is genuinely the impressive part here. But this one experiment doesn't prove AI can build any kind of organism on its own, big or small. What it really shows is people and AI working together, not AI working alone.
Evo 2 is free for anyone to download and use. Stanford made that choice on purpose. The idea is that more researchers using the tool means faster progress on real biology problems. But that same openness comes with a downside. If anyone can access a tool this powerful there's no way to fully control who uses it or why.
Brian Hie, the researcher behind the project, acknowledges the risk. Stanford reports that modified versions of Evo 2 could potentially be misused by bad actors. However, Hie argues that open access can speed up legitimate research and that AI tools could also help scientists respond to natural pandemics and man-made biological threats.
It helps to keep this in perspective though. Having access to Evo 2 does not mean someone could just design a dangerous virus and release it. Building an actual working virus still requires synthesizing real DNA and running lab tests, steps this model can't do on its own. And this particular experiment only involved viruses that attack bacteria, not humans.
Researchers want to test genome generation on other bacteriophages and on more complex DNA, while working to improve how novel and controllable the generated results can be. This experiment is described as a starting point rather than a finished capability.
Stanford says Hie is working with other researchers on additional bacteriophages and is also investigating longer, more complex DNA sequences. There is a possibility of eventually exploring small bacterial genomes. Looking further ahead, researchers have discussed possible future phage work involving harmful bacteria such as MRSA, Pseudomonas aeruginosa, and the bacteria linked to tuberculosis. Hie points to greater genetic novelty and greater control over outcomes as the major open questions still facing this line of research. These are future directions, not capabilities the current experiment has already demonstrated.
The core distinction from this study still holds as the field moves forward. Genome models can now contribute working genetic designs, but real computational and physical validation work still separates an AI-generated sequence from a useful biological system.
Sort of, but not on its own. Evo 2's job was to design a complete virus genome, which it did successfully. However, a digital design isn't a real virus yet. Researchers still had to manufacture the actual DNA, get it into bacteria, and run lab tests to see if it functioned. Sixteen designs ended up working. So while AI handled the creative design part, humans handled every physical step that turned that design into something real.
Think of Evo 2 as a system trained to read and write DNA, similar to how AI chatbots are trained to read and write language. It studies real genetic sequences and learns their patterns. Then, starting from a small piece of real DNA, it can generate an entirely new, complete sequence on its own. In this study, researchers used it to generate whole virus genomes based on a real bacteria-infecting virus.
Why does the article mention two different AI models?
The research actually involved two versions of Stanford's AI, Evo 1 and an updated version called Evo 2. Both were used at different points during the project, but Stanford's public announcement focuses mainly on Evo 2 as the highlight. The exact breakdown of which model produced which specific results hasn't been fully detailed yet, so that distinction is something worth watching once the full scientific paper gets published.
In theory, maybe, but not practically. Manually designing a full, working genome without AI assistance would take an enormous amount of time, expertise, and trial and error. That's exactly the kind of complex pattern recognition AI tools are good at handling. The technology didn't replace human expertise here, but it dramatically sped up a process that would otherwise be extremely slow and difficult to do by hand. (76 words)
Phage therapy isn't a new concept, scientists have studied it for decades in some parts of the world. What's changed recently is the technology available to design and refine phages more precisely. AI tools now make it possible to generate custom genome designs faster than researchers could manage manually, which could help revive interest in phage therapy as antibiotic resistance becomes a bigger global concern. (76 words)
