Recently, I have been thinking about how people in the AI safety community view societal impacts research. Broadly, this work aims at exploring how AI is used and affects the real world.[1] I think there is a misunderstanding about what societal impacts research actually aims to accomplish. I believe that people conflate "societal impacts research" with "long timeline thinking," so short timeline people devalue it, when it actually applies regardless of timeline.
I think it is important to be precise about what "short timelines" mean, as it can point to two different worlds. One is a world where AI goes well and transforms society fast and the other is a world where it does not go well. The argument I am making here is squarely about the first. Thus, the case for societal impacts research holds regardless of whether that world arrives in three months or thirty years. If timelines are short and AI does go well, the case for societal impact research only strengthens, as there is less time to adapt.
Now, I have heard it said by short timeline people that if AI does go well, then societal impacts research will naturally follow. While I do think this is a path, I am doubtful about whether this is the most effective path. And by leaving it to naturally flow, it is easy for someone like me to perceive this as a lack of concern or an abdication of effort to steer it.
I'm not asking for short timeline people to give up their view of the world. What I am asking is that they consider societal impacts research as timeline agnostic and something that should happen concurrently with development. I believe that efforts should be made to increase the opportunities to do societal impacts research, in the form of fellowships, residencies, or grants. Besides, the marginal cost of supporting it now is low next to the incalculable cost of picking up the pieces after the fact.
Recently, I have been thinking about how people in the AI safety community view societal impacts research. Broadly, this work aims at exploring how AI is used and affects the real world.[1] I think there is a misunderstanding about what societal impacts research actually aims to accomplish. I believe that people conflate "societal impacts research" with "long timeline thinking," so short timeline people devalue it, when it actually applies regardless of timeline.
I think it is important to be precise about what "short timelines" mean, as it can point to two different worlds. One is a world where AI goes well and transforms society fast and the other is a world where it does not go well. The argument I am making here is squarely about the first. Thus, the case for societal impacts research holds regardless of whether that world arrives in three months or thirty years. If timelines are short and AI does go well, the case for societal impact research only strengthens, as there is less time to adapt.
Now, I have heard it said by short timeline people that if AI does go well, then societal impacts research will naturally follow. While I do think this is a path, I am doubtful about whether this is the most effective path. And by leaving it to naturally flow, it is easy for someone like me to perceive this as a lack of concern or an abdication of effort to steer it.
I'm not asking for short timeline people to give up their view of the world. What I am asking is that they consider societal impacts research as timeline agnostic and something that should happen concurrently with development. I believe that efforts should be made to increase the opportunities to do societal impacts research, in the form of fellowships, residencies, or grants. Besides, the marginal cost of supporting it now is low next to the incalculable cost of picking up the pieces after the fact.