I think there's a lot of misunderstanding about AI causing dangerous runaway biorisk. Here's what I said to a worried colleague: I think there are three aspects of this that are important. Firstly, there's lots of things in biology that are just unknown and for which we don't have the data, and the attrition rate is so high that neither humans nor AI can do very well. Drug discovery is a spectacular example - most things just don't work, and we are always swimming in a sea of ignorance. Secondly, lab work is extremely messy and prone to false positives and negatives and things just not working. AI can help accelerate some of this work - I have been part of such projects myself - but we are very far from AI being able to autonomously orchestrate any kind of risky work in a way that poses unknown dangers, except those that would arise from accidents (glassware and robotic arms breaking, chemicals being mixed unintentionally). Thirdly, all the autonomous biological discovery that I am seeing is almost exclusively computational and falls into two categories: discovery with a human strongly in the loop, and discoveries where AI has sped up an already well-understood process. That being said, I agree that it's a good idea to sandbox or gate AI for biology or chemistry experiments with a strong human-in-the-loop component, but more so in my mind because it could spend valuable capital chasing minor goals or cause dumb accidents rather than design a supervirus autonomously or anything like that. I am also in favor of pausing new model development for, say, a year while we discuss future potential biorisk and agree on international safeguards, the way it was done at Asilomar with recombinant DNA technology in the 70s. I think the bigger issue in some sense for me from an optics standpoint is that if tomorrow a bad actor designs a gain-of-function mutation "using AI", even if it just means speeding up an existing process or doing literature searches faster or better - basically stuff they could have done without AI - the alarmists are going to yell, "We told you so!"