AI has left the screen

For most people, the consequences of generative AI still arrive through a screen. A model produces a paragraph, an image, a song or a piece of software. The output may be useful, misleading or harmful, but it remains information.

This month, that boundary moved.

Researchers from Stanford University and the Arc Institute used genome language models to generate complete DNA sequences for bacteriophages, viruses that infect bacteria. Scientists synthesized nearly 300 of the proposed genomes and tested them against Escherichia coli. Sixteen produced viable phages. Some performed better than the natural virus used as their starting point, and a mixture of the generated phages killed bacteria that had developed resistance to that natural virus.

The result, published in Science on August 6, is easy to sensationalize. AI did not independently create life from nothing. Researchers selected the target, built the design framework, filtered the candidates, paid to synthesize the DNA and conducted the laboratory experiments. The viruses are bacteriophages, not pathogens capable of infecting people. Even calling a virus “life” enters a scientific debate that the experiment does not resolve.

None of those qualifications erase what happened.

An AI system generated complete genomes. Human researchers turned a fraction of those digital sequences into physical viruses. Those viruses infected bacterial cells, reproduced through them and killed bacteria in the laboratory.

The important threshold is not autonomy. It is translation.

From reading DNA to writing it

Genome language models operate on a principle that will sound familiar to anyone who has used a chatbot. A conventional language model learns statistical relationships between words and predicts what should come next. A genome language model learns patterns across DNA and predicts the next nucleotide.

The similarity ends quickly. A plausible sentence only has to resemble language. A viable genome has to satisfy thousands of interacting biological constraints well enough to function inside a cell.

The model used in the experiment, Evo 2, was developed by researchers from the Arc Institute, Stanford, NVIDIA, the University of California, Berkeley and other institutions. According to its peer-reviewed description in Nature, the largest version contains 40 billion parameters and was trained on more than nine trillion nucleotides. Its training data spans bacteria, archaea, plants, animals, humans and bacteriophages. Viruses that infect humans and other eukaryotic organisms were deliberately excluded as a safety measure.

For the bacteriophage experiment, researchers worked from ΦX174, a well-studied virus with a genome of roughly 5,400 nucleotides that infects E. coli. The system generated possible genomes, but the path from possibility to biology remained intensely human. Researchers evaluated the sequences against design criteria, chose which ones justified the cost of synthesis and tested the resulting material in the laboratory.

Only 16 of the nearly 300 synthesized designs proved viable. That success rate is both a limitation and an achievement.

It is a limitation because the model did not master biology. Most of the physical designs failed. It is an achievement because biological evolution ordinarily does the filtering through mutation, reproduction and death. Here, a model proposed genetic possibilities that natural history had apparently never tried, and some of them worked on the first experimental pass.

The gain is not hypothetical, but it is preclinical

The medical argument begins with antimicrobial resistance.

Bacteria evolve. When antibiotics are overused or misused, the organisms they fail to kill can survive and pass their resistance forward. The World Health Organization estimates that bacterial antimicrobial resistance was associated with more than 4.7 million deaths globally in 2021.

Bacteriophages offer another way to attack those infections. Unlike broad-spectrum antibiotics, a phage can be highly specific to a particular bacterium or even a particular bacterial strain. It infects the bacterial cell, uses that cell to reproduce and eventually destroys it.

That specificity is the attraction and the problem. A phage that kills one strain may do nothing against another. Bacteria can also evolve resistance to a phage just as they evolve resistance to an antibiotic.

The Stanford-led experiment points toward a possible response: generate several genetically different phages and combine them. A bacterium might escape one member of the cocktail, but escaping all of them at once should be harder. In the laboratory, the researchers reported that their 16-phage mixture overcame E. coli that had become resistant to the natural ΦX174 phage.

That is a measured gain. It is not yet a medicine.

The work involved bacterial cultures, not patients. The generated phages have not been shown to treat an infection in an animal or human. Researchers still need to examine toxicity, immune response, dosage, delivery, manufacturing consistency and the possibility that bacteria will develop new forms of resistance.

Phage therapy itself is not new. Physicians have experimented with it for more than a century, and difficult infections have sometimes been treated through clinical trials or compassionate-use programs. Yet its clinical adoption remains limited by narrow host ranges, inconsistent outcomes, manufacturing requirements and regulation. In the United States, unlicensed phage products used in patients require an Investigational New Drug application or expanded-access authorization from the Food and Drug Administration.

AI may accelerate the design step. It does not remove the rest of medicine.

That distinction matters because a separate August review in Nature Reviews Drug Discovery reached a sobering conclusion about the larger field. After a decade of AI methods, benchmarks and announcements, evidence of clinically relevant impact on producing safer and more effective medicines faster remains limited.

The 16 bacteriophages are therefore significant for what they demonstrate, not for what they have already delivered. AI-generated genomes can function. Whether that capability becomes a reliable treatment is still an open experiment.

The same capability has another direction

There is no honest way to discuss a system that writes viral genomes without discussing misuse.

The experiment itself was deliberately confined to bacteriophages. Evo 2’s developers excluded viruses that infect humans, animals, plants and other eukaryotic organisms from its training data. Their safety evaluations found that the model performed poorly on human viral sequences, suggesting the exclusion meaningfully reduced its ability to work with human pathogens.

That is evidence of a safeguard. It is not proof of a permanent boundary.

A March 2026 preprint tested whether Evo 2’s excluded capabilities could be partially restored through additional training. Researchers fine-tuned the open-weight model using sequences from 110 harmful human-infecting viruses. The modified model improved on some predictive tasks involving unseen viral sequences. The study did not generate a human pathogen, and it has not completed peer review. It nevertheless shows why removing dangerous material from the original training set may not be sufficient once a model’s weights can be downloaded and modified.

This is the central tension of open biological models. Open access can distribute scientific capacity beyond a small number of companies and elite laboratories. It can help researchers study rare diseases, engineer useful microbes and respond to emerging pathogens. It can also make a safety decision taken by the original developer easier for another user to reverse.

The final barrier between a digital genome and a functioning organism is physical synthesis. That barrier is real. DNA synthesis costs money, laboratory work requires skill and equipment, and many generated sequences will fail. Screening systems can also flag concerning orders and verify whether a customer has a legitimate reason to request them.

Current safeguards, however, are not universal. U.S. guidance recommends screening synthetic DNA and RNA orders for sequences associated with pathogenicity or toxicity, verifying customers and applying safeguards to benchtop synthesis devices. The 2024 federal framework tied compliance to certain federally funded life-sciences research, and the government has since directed agencies to revise or replace it. These measures do not yet create one comprehensive global system covering every model, synthesis provider, laboratory and researcher.

The World Health Organization’s guidance on dual-use life-sciences research treats responsibility as something shared across the entire research cycle. That includes model developers, scientists, universities, funders, DNA synthesis providers, publishers and governments. Genome-generating AI makes that shared model more necessary because no single checkpoint controls the entire process.

What the experiment actually changes

It would be easy to end with either of two headlines.

AI designed viruses that could help defeat antibiotic-resistant bacteria.

AI designed viruses before governments were prepared to govern the capability.

Both are supported by the record. Neither is complete by itself.

The experiment did not produce a new antibiotic, but it produced a path worth testing. It did not create a human pathogen, but it demonstrated a capability that safety systems will now have to follow. It did not remove scientists from the process, but it changed what scientists can attempt.

Until now, much of the argument over generative AI has concerned whether a model can imitate human output. Can it write like us, draw like us or reason like us?

Biology introduces a different test. The genome does not care whether its sequence sounds convincing. It either functions or it does not.

Sixteen of these did.

That is why August’s development deserves more than wonder or fear. It deserves a trace. The gain is the possibility of designing new tools against organisms that have learned to survive our existing medicines. The problem is that the same design capacity does not arrive with a guaranteed purpose.

AI has not become a biologist working alone. It has become a new instrument inside biology. What matters now is who uses it, what they ask it to design and which outputs are allowed to leave the screen.

Sources cited in this reading

Every factual claim in this reading is tied to a peer-reviewed publication, government source or clearly identified preprint.

• Samuel H. King et al., “Generative design of bacteriophages with genome language models”, Science, August 6, 2026.
• Thomas V. Inglesby and Moritz S. Hanke, “AI-designed viral genomes”, Science, August 6, 2026.
• Garyk Brixi et al., “Genome modelling and design across all domains of life with Evo 2”, Nature, March 2026.
• Andreas Bender et al., “Artificial intelligence in drug discovery: what it is, where we stand and the path forward”, Nature Reviews Drug Discovery, August 7, 2026.
• World Health Organization, “Antimicrobial resistance”, updated July 16, 2026.
• U.S. Department of Health and Human Services, “Synthetic Nucleic Acid Screening”.
• World Health Organization, “Global guidance framework for the responsible use of the life sciences”, September 2022.
• James R. M. Black et al., “Open-weight genome language model safeguards: Assessing robustness via adversarial fine-tuning”, preprint revised March 2026.
• U.S. Food and Drug Administration, “Science and Regulation of Bacteriophage Therapy”, 2021.