Q: One of the dangers of AI is that is may be more easy to make a new dangerous lethal virus. However, how easy is that really? It may be hard to predict how a new virus will impact people and also a more lethal virus will have more difficulty to spread widely (as people die too soon to infect other people). A non lethal virus may be a nuisance, but will not eradicate human kind. What is your take?
A: I think the risk is real, but sometimes described too dramatically. There is a very large gap between “AI knows a great deal about viruses” and “AI can enable someone to create a virus that wipes out humanity.”
Current frontier models already show roughly expert-level performance on some advanced biology questions, which is why governments are taking the biosecurity issue seriously. But biology has an important property that software does not: you cannot reliably predict from the design what will happen in the real world.
I would see four increasingly difficult steps:
1. AI helps someone understand or design biology — increasingly easy.
AI can search enormous literatures, analyse biological data and help experts formulate hypotheses. This undoubtedly lowers some barriers.
2. Producing a viable new pathogen — much harder.
You still need biological materials, laboratory facilities, expertise, experimentation and repeated testing. Many apparently clever designs simply won’t work.
3. Creating something that causes a global pandemic — much harder again.
It would simultaneously need properties such as efficient human-to-human transmission, sufficient stability, successful replication in humans and avoidance of existing immunity. These characteristics interact in complicated and often unpredictable ways.
4. Creating something capable of human extinction — an extraordinarily high bar.
Humans differ genetically and immunologically; populations are geographically dispersed; some people would probably be naturally resistant; behaviour would change; and vaccines, antivirals, isolation and other countermeasures would begin once the threat was detected.
And you are partly right about lethality. Killing or incapacitating a host quickly can reduce the opportunity to transmit, which is the classic transmission–virulence trade-off. But it isn’t a biological law. Recent reviews stress that the relationship is much more complicated.
The critical variable is timing. Imagine a hypothetical disease where someone:
becomes infectious → feels perfectly healthy for ten days → infects many people → then becomes seriously ill.
High mortality at the end would do little to prevent transmission that had already occurred. Conversely, a disease that makes someone extremely sick before they become infectious would be much easier to contain.
So I would place the risks roughly like this:
AI-assisted biological accidents or deliberate outbreaks: credible concern.
AI contributing to a severe future pandemic: plausible enough to justify safeguards.
AI enabling an amateur effortlessly to create a civilization-ending virus: not where the technology is today.
A virus exterminating humanity: conceivable in abstract risk analysis, but biologically much harder than the phrase “AI can design a deadly virus” suggests.
There is also an important counterweight that gets much less publicity: the same AI capability can accelerate outbreak detection, diagnostics, antiviral development and vaccine design. NTI explicitly describes both sides of this convergence. The Nuclear Threat Initiative
This actually connects rather neatly to our earlier discussion. Intelligence is not magic. An AI becoming extraordinarily knowledgeable about biology does not mean that biology suddenly becomes completely predictable or controllable. Living systems remain messy, evolutionary and full of surprises.
And that may be one of the more reassuring limits on extreme AI bio-risk.
