AI designed hundreds of viruses. 16 worked, and that's fine

A quiet post about a loud result

On August 7, Arc Institute posted something that reads like lab housekeeping. One of its most popular preprints of 2025, @arcinstitute wrote, "is now published in @ScienceMagazine," with congratulations to Brian Hie, Samuel King and the team for a framework for "generating and evaluating AI-generated genomes."

That preprint first went up in September 2025, when @BrianHie announced it with "Welcome to the age of generative genome design!" and a nod to history: the first genome ever sequenced, back in 1977, was the bacteriophage ΦX174. His lab used that same phage as its design template.

Peer review took eleven months. The paper ran in Science on August 6, and this time the story did not stay inside biology. @Forbes pushed it to a general audience with a headline about an AI model inventing 16 new viruses, and by that evening it was in the regular news cycle.

What actually happened

Researchers at Stanford and Arc Institute used two genome language models, Evo 1 and Evo 2, to write complete viral genomes from scratch. They filtered the output computationally, sent close to 300 designs out to be synthesized as physical DNA, and tested what came back against E. coli.

Sixteen of them worked.

It is worth being precise about the word viruses here. These are bacteriophages, viruses that infect bacteria and cannot infect people. Evo's training data excludes human viral sequences, and the experiments ran on non-pathogenic lab strains of E. coli inside biosafety cabinets.

The results are real science, not a demo. Cryo-electron microscopy showed one generated phage packaging its DNA with a protein borrowed from an evolutionarily distant relative, a swap that earlier engineering attempts had failed to pull off. A cocktail of the generated phages then cleared E. coli strains that had already evolved resistance to ΦX174. That is the entire promise of phage therapy against antibiotic-resistant infections, shown working on a bench.

The number nobody puts in the headline

About 5% of the designs worked.

Simon Jackson, a phage therapy researcher quoted by the Science Media Centre, put the hit rate at roughly one in twenty and called the work an impressive proof of principle. Jordi García Ojalvo, a systems biology professor at Pompeu Fabra University in Barcelona, was blunter about the misses, using the word hallucinations. Same failure mode you get from a chatbot. The difference is that a wet lab settles the argument instead of a reader.

That is the part I keep coming back to. The model was wrong roughly 95% of the time and the result still counts as a landmark, because the people running it had a cheap, objective test they could apply hundreds of times. The assay decided. The model only proposed.

The oversight gap, stated plainly

Science ran a companion commentary the same day from Thomas Inglesby and Moritz Hanke at the Johns Hopkins Center for Health Security. Their summary has been quoted everywhere since: the ability to compose viral genomes with generative AI now exists, and the governance to steer it safely does not. They want DNA synthesis providers legally required to verify both the sequence they are printing and the identity of whoever ordered it. Neither requirement is in US law today.

Reasonable people can disagree about how urgent that is for phages specifically. Nobody serious disagrees that the check has to exist before the capability scales to bigger genomes.

What a business should take from this

None of this touches your operations. The shape of it does.

Every AI deployment I have seen actually work has two parts. There is a generator that is fast, cheap and wrong a lot. And there is a filter that is slower, boring, and right. Stanford had a growth inhibition assay. Your version might be a reconciliation against the accounting system, a person reading the top five drafts before anything sends, or a rule that no quote over a certain dollar amount goes out without a signature.

Most small businesses wiring up AI right now have built the generator and skipped the filter. The output reads well, so it ships. Then a customer gets a wrong price or a follow-up email meant for someone else, and the cleanup costs more than the tool ever saved.

So the question to answer before automating anything: what is my assay? If you cannot state the pass or fail test in one sentence, the automation is not ready, however good the drafts look.

We run a free process audit for businesses around the Fox Valley and across Illinois, and that is usually the first question on the table. Sixteen out of nearly three hundred is a terrible batting average anywhere except inside a system built to throw the rest away.