In the near term, the biggest worry isn’t so much AI going rogue, but being used to create biological weapons and carry out cyberattacks. This is what Bill Gates is most worried about. Anthropic also believes that biological misuse is one of the most serious risks of frontier AI models.
It’s good that Gates, Anthropic and other AI companies are highlighting those risks, but I agree with Ash Jogalekar, who points out that the evidence for AI increasing biological threat capabilities in the short term is limited.
Biological weapon design and drug design are subject to similar constraints. From experience, I know that AI is already making drug target discovery research easier. When coupled with the ability to inform lab experiments, it can accelerate certain aspects of drug discovery, especially during early development. However, these experiments still need to happen, and they still take time. In other words, using AI for useful (or dangerous) research is different from making an effective drug (or a viable biological weapon). Here is how Jogalekar puts it:
There is no doubt at this point that AI gives you access to an incredible amount of compressed and potentially actionable information.
But the jump from that to real-world biological capability is much larger than it is often made to sound. […] The problem is not a lack of models, it’s simply what we don’t know. The value of ignorance in biology cannot be underestimated.
Especially in biology, the downstream steps are often where most of the science lives.
One surprising reason for AI’s limitations in drug discovery is that they’re trained on the scientific literature. The number of papers on a drug target is not predictive of a drug making it from early development to FDA approval. The deeper reason is that biology is complex. The 40 million biomedical papers that have been published to date are only scratching the surface.
That’s why I don’t think that AI will lead to a large increase in approved drugs in the next few years, and why I think that in the short term, AI biosecurity risk will remain manageable.
