AI Is Playing God

Artificial intelligence has learned to write, code, make images, and even generate videos. Now, scientists have taught it something considerably more unusual: how to design viruses.

That sounds like the opening scene of a sci-fi movie, but researchers at Stanford University and the Arc Institute have demonstrated that AI can learn the “grammar” of DNA and use it to design viral genomes that actually work.

The experiment could eventually lead to new medical tools — while also raising uncomfortable questions about how far AI-driven biology should go.

AI Just Learned How To Write Viral DNA

The breakthrough centers on an AI model called Evo, developed to understand DNA much as language models learn patterns in text.

Instead of training Evo on books and websites, researchers trained it on genetic sequences from millions of organisms, including animals, plants, microbes and viruses. In total, Evo analyzed roughly 9 trillion DNA nucleotides, the molecular building blocks that make up genetic material.

The idea was simple: DNA has rules. Certain sequences work together, while others produce biological nonsense. If AI can learn those rules, perhaps it can generate new DNA sequences that biology can actually understand.

Researchers first used Evo to design new genes. Then they decided to take things a step further.

Scientists Asked AI To Build A Virus

The team focused on Phi X-174, a bacteriophage, or virus that infects bacteria rather than humans. Scientists have studied the virus for nearly a century, making it a relatively well-understood and contained starting point.

Researchers trained Evo on the 11 genes of Phi X-174 and roughly 15,000 related viral genomes. They then asked the AI to generate new versions.

Evo identified about 700,000 potential viral genomes.

Scientists selected 285 of those designs for laboratory testing. They synthesized the corresponding DNA and inserted it into E. coli bacteria to see whether the genetic instructions could actually produce viruses.

They could.

Sixteen of the AI-generated genomes produced viable viruses that could infect bacteria. Some even multiplied faster than the natural Phi X-174 virus.

That is the part that makes the experiment particularly significant. Evo wasn't simply producing DNA sequences that looked convincing on a computer. Its designs worked in the real world.

Why AI-Designed Viruses Could Help Medicine

Before anyone imagines a robot creating the next pandemic, there is an important distinction: These viruses were designed to target bacteria, not humans.

And that could eventually be useful.

Bacteriophages naturally attack bacteria, and scientists are increasingly studying them as potential alternatives or complements to traditional antibiotics. That is particularly interesting as antibiotic-resistant infections become harder to treat.

Researchers are already exploring “personalized” phage therapies, where a virus is selected or engineered to attack a particular bacterium infecting a patient.

AI could eventually accelerate that process by helping scientists search through enormous numbers of possible viral designs much faster than humans could.

In other words, the same technology that sounds unsettling could potentially help researchers build more precise weapons against dangerous bacteria.

But Here's Where Things Get Uncomfortable

The researchers were careful about what Evo could learn.

They deliberately excluded genetic information from viruses that infect humans, as well as related viruses that infect animals, plants and fungi. That means the model used in this experiment wasn't designed to generate human pathogens.

But the broader possibility is what has scientists and biosecurity experts worried.

If AI models become increasingly capable of understanding biological systems, they could potentially lower the barrier to designing biological agents that are more difficult to create using traditional methods.

That creates a tricky problem for regulators: How do you determine whether something is dangerous when it has never existed before?

A naturally occurring virus can be identified and classified. An AI-designed virus could be something entirely new.

AI Biology Is Moving Fast

The study highlights a much bigger shift happening at the intersection of AI and biology.

For years, AI has mostly been used to analyze biological information — predicting protein structures, identifying drug candidates and helping scientists understand diseases.

Now, researchers are increasingly asking AI to design biological systems.

That's a much bigger leap.

In this experiment, AI didn't replace the scientists. Humans chose the virus family, trained the model, selected which designs to test, and conducted the laboratory experiments.

But the ability to generate hundreds of thousands of biological designs in a matter of time could dramatically expand what researchers can test.

And that is both the promise and the problem.

AI may eventually help scientists develop new antibiotics, therapies and biotechnology tools faster than ever. But as these models become better at designing biology, researchers and governments will also have to figure out where the guardrails need to go.

For now, AI has learned to make viruses that infect bacteria.

The bigger question is what happens when AI gets much better at understanding the biology of everything else.

That's all for this week! If you found these updates useful, you'll like more from this newsletter. Get deeper dives, hot takes, and all the latest tech news delivered straight to your inbox.