Many of you readers don’t spend your time on twitter; good for you! But I wanted to share this quick thread, since it seems like lots of people are confused about the existence of expert agreement about how much current AI enables bioterrorism. Below is the thread. (Also on BlueSky!)
Hot take: Most AI-bioterrorism risk "debate" is people shadow-boxing over definitions, not different predictions about concrete points. I've talked to many other experts. We often don’t agree on what “enable” means, so people think the risk is uncertain in ways it is not.
1/ In early 2024 no LLM could walk a garage bioterrorist past many different hard wet-lab hurdles. They still needed experience with protocols and tacit know-how, plus access to various materials that they wouldn't know how to get.
2/ Everyone agrees that 2025 models shrink that gap a lot: they can provide clever ideas for how to get access to key materials, have better ability to explain how to proceed, and do better troubleshooting. A motivated but middling lab jockey would now get further, faster.
3/ But now, we get into lots of terminology and confusion. What is high risk? What is capability? What are people going to try to do with the models? Will the models be monitored? Will the models refuse? Will this be reported? What counts as failing?
4/ Is a model “high-risk” only if it can hand-hold you from zero to aerosolized pathogen, end-to-end? (The current definition used for the OpenAI and DeepMind high risk threshold.) What if it can do all but one step? And what if people might figure that one out on their own?
5/ Is shaving 30% off the time needed for each critical step—design, synthesis planning, scale-up—already a dangerous capability jump? What if it makes it 30% more likely to succeed? Does it matter if it's moving from 90% failure rate to 60%? Or from 60% to 30%? Or 30% to 0%?
6/ Then there’s the refusal shell-game: If a model is otherwise capable, but only answers 5% of disallowed queries, is it “incapable” or merely poorly supervised? (And are bioterrorists likely to be good at prompt engineering? Will they get better, so refusals are ~irrelevant?)
7/ Next, marginal-risk math: if each new model adds, say, 3% to the risk, and there's a new model every month (if only it were that slow!) that’s 40% more risk in a year - but each provider says their model isn't adding much risk. Which is true, but still, risk is now 40% higher.
8/ And threat actors differ. Lone cranks, cults, and state proxies each have different time frames, risk tolerance, and funding. So, does it count as "enabling" if a group could do it before, but it will now cost 10% as much, and happen 5x faster?
9/ And models are reducing the failure rate of multiple steps, which has a cumulative impact on overall failure rate. Speaking of which, speed matters. Fewer trial-and-error loops means fewer red flags for intel and fewer chances wannabe bioterrorists bail out in frustration.
10/ Lastly, no one knows how much individuals and terrorist groups will shift efforts from conventional attacks to chemical or bio attacks, if at all. And we have very little information about how often groups try and failed, much less considered it and didn't try, in the past.
11/ So, does it count as disagreeing about whether models are high risk if we don't agree on these terms? Yes, but the disagreement is verbal, not substantive. But everyone agrees that the models can enable far more dangerous uses, and make attempts more likely to succeed.
12/ Stop with reductive binary questions like "does this count as high risk under industry-suggested definitions." Instead, discuss “capability”, “enablement”, and “risk increment” explicitly, and then it will be clear we're agreeing the risk keeps trending upward.
13/ To finish, the above is explanation and clarification of expert agreement about *today's LLMs*. And there are uncertain but definitely far larger risks from newer AI models. We don't know how quickly risks will increase, but we can see that no-one is ready to address them.

