AI-assisted design of novel viruses poses less public-health risk than widely feared

What's this about?

People disagree about whether AI help to design new viruses creates less danger than many people fear.

The key issue is whether a computer plan can become a virus that harms and spreads among people.

What supporters say

• An AI-made gene plan is only the first step toward making a real virus.
• A team still needs cells, lab tools, skill, and many tests to make a virus work.
• A virus must infect people, spread well, and cause harm before it can spark a big health threat.
• Tests have made new viruses that infect germs, but those viruses do not infect people.

What critics say

• The danger is real because AI may help some people create virus plans faster.
• A harmful virus could become more likely if lab skills and tools become easier to get.
• Small changes through many rounds of tests could help a virus work better over time.
• Health groups and DNA sellers need strong checks before digital plans become real germs.

The bottom line

AI-made virus plans do not yet equal viruses that can cause a major outbreak.

So the near-term risk seems lower than the scariest claims, but the risk could grow.

AI-assisted design of new viruses may pose less immediate public-health risk than the most alarming accounts suggest, but the danger is real and could grow. The key distinction is between generating a genetic sequence on a computer and creating a pathogen that can infect, spread among and seriously harm people.

The case for

The strongest argument for near-term reassurance is that AI-generated sequences are only the first step in a long and difficult process. A computer model may produce a plausible viral genome, but that does not mean the virus will work in a cell, cause disease, spread efficiently or survive outside a laboratory. Researchers would still need the right cell systems, laboratory methods, technical skill and repeated testing to turn a design into a functioning pathogen.

RAND’s analysis describes several hurdles beyond sequence generation: biological viability, the traits of the virus, delivery, laboratory execution and further rounds of improvement. In other words, an AI-produced genome is not the same thing as an outbreak-capable virus. These bottlenecks substantially limit immediate risk 1.

The clearest experimental proof of AI-assisted viral design also has limited direct relevance to human health. In a peer-reviewed Science study, researchers used a genome language model to generate new bacteriophage genomes. Some selected designs were synthesized and shown to work experimentally. But bacteriophages infect bacteria, not people. The work does not show that AI can create a human virus with the combination of transmission, immune evasion, severity and environmental stability needed to cause a major public-health crisis 2 (see Figure 2).

There are also points at which authorities and companies can intervene before a digital design becomes a biological reality. U.S. guidance recommends screening orders for synthetic genetic material and checking customers. World Health Organization guidance points to other layers of protection, including laboratory containment, trustworthy personnel, inventory controls and plans for responding to incidents. These measures are imperfect, but they offer practical barriers between software and a physical virus 3.

The case against

The concern cannot be brushed aside as hypothetical. The phage studies show that AI models can learn biologically meaningful features of viral genomes and generate designs that function in experiments. While this is not evidence of human-pathogen design, it is evidence that the basic capability is advancing 4.

Risk could rise if AI becomes useful across more parts of pathogen development, rather than merely suggesting sequences. AI might help researchers move faster through design and testing cycles, or make specialized knowledge easier to use for people with less training. That would matter most when paired with accessible laboratory equipment and the ability to run repeated experiments. Current models still face major limits, and physical infrastructure remains essential, but AI could lower the expertise and iteration costs over time 5.

Existing safeguards also have gaps. Screening systems do not cover every provider, standards vary, and highly novel or altered sequences can be difficult to identify as dangerous. Systems designed to compare orders with known threat sequences may struggle when a new design looks sufficiently different from what is already catalogued. False alarms, missed threats and limited data for validating more advanced “function-aware” screening remain unresolved problems 6 (see Figure 3).

Perhaps the biggest unknown is how often AI-designed viruses are actually attempted, successfully built or used in ways that cause harm. There is no reliable estimate of real-world harmful use involving AI-assisted viral design. Much of the debate about future threats rests on reviews, policy analysis and plausible scenarios rather than documented malicious deployment or public-health outcomes. Unresolved questions about conflicts of interest in parts of the literature add another reason for caution when weighing the evidence.

The bottom line

The evidence supports a qualified, high-confidence conclusion: AI-assisted viral design presents a nontrivial risk, but its near-term public-health danger is constrained by major biological, technical and operational hurdles.

The claim is defensible only in a narrow sense. It is reasonable to say the immediate risk is lower than accounts that treat generating a viral sequence as equivalent to creating an outbreak. It is not reasonable to conclude that the threat is insignificant, globally controlled or unworthy of precaution. Risk depends heavily on the target organism, the model’s real-world capabilities, the user’s access to laboratories and the strength of local screening and biosecurity systems.