Thanks for this, Vael! As I said previously, here are some areas of agreement and potential disagreement.
Agreement
I generally agree that most people with social science PhDs should look outside of AI safety research, and I like your suggestions for where to look.
I think that the top 10% or so of social science researchers should probably try to do AI safety related research, particularly people who thrive in academic settings or lack movement building skills.
Overall, I’d encourage anyone who was equally good at AI safety SS research and AI safety movement building, to choose the latter option. AI safety movement building probably feels higher expected impact to me than AI safety SS research because I think that movement building is the highest impact ‘instrumental’ cause area, at least until vastly more people know about, understand and work on the key concepts, arguments and needs of EA.
Potential disagreement (shared as a total non-expert, to be clear)
My intuition is that AI safety research is still relatively undersupplied by social science researchers compared to the ideal. I think the area could, and ideally should, absorb a lot of social science research people over the next 30 years if funding and interest scale as I expect. Maybe 5,000+, if I consider all the organisations and geographies involved. Ideally, as many as possible of these people would be EA aware and aligned.
What might this look like? For instance, in academia, I see work to i) understand if/where creating evidence and interventions might be useful (e.g, interviewing/surveying technical researchers, policymakers and organisational leaders and/or mapping their key behaviours to influences), to ii) prioritise what is important (e.g., ranking malleability of different interventions/behaviour), then doing related research. I also foresee a range of theoretical work around how we can port over concepts and theory from areas such as communication, psychology and sociology to describe, understand and optimise how human and machines interact. I also expect that there will be a lot of value in coordination to support that work. I expect to see a global distribution of research labs researcher and projects.
In government and private settings, I foresee social science researchers hired to do more ‘context bound’ work with clear connections to immediate policy decisions. Government embedded research teams who need to understand technical, political and social factors to understand how to craft effective different types of national policy. Organisationally embedded teams engaged to create organisational policies that get value from internal AI, or engagement with other organisations platforms. Lots of work in the military and defence sector.
In support of these areas, I see a lot of social science background people being useful by working as knowledge brokers to effectively translate and communicate ideas between researcher and different types of practitioners (e.g., as marketers, community builders, user researchers, or educator). Also in providing various research support structures, training and curating potential students and research assistants, setting up support infrastructure like panels of potential technical/policy research participants, starting/organising conferences and journals etc).
I also suspect that’s there’s going to be lots needed that I have left out.
Overall, from a behavioural science research perspective (e.g., who needs to do what differently/what is the most important behaviour here/how do we ensure that behaviour happens), there are a lot of different behaviours, audiences contexts, and interactions, and little to no understanding of the key behaviours, actors, or ideal interventions. If this is life-on-earth-threatening-in-the-near-future stuff, then there is a lot of work to be done across a huge range of areas!
Not sure if this is an actual disagreement, so let me know. It’s useful for me to write up and share regardless, as it underpins some of my movement building plans. Feedback is welcome.
Thanks for this, Vael! As I said previously, here are some areas of agreement and potential disagreement.
Agreement
I generally agree that most people with social science PhDs should look outside of AI safety research, and I like your suggestions for where to look.
I think that the top 10% or so of social science researchers should probably try to do AI safety related research, particularly people who thrive in academic settings or lack movement building skills.
Overall, I’d encourage anyone who was equally good at AI safety SS research and AI safety movement building, to choose the latter option. AI safety movement building probably feels higher expected impact to me than AI safety SS research because I think that movement building is the highest impact ‘instrumental’ cause area, at least until vastly more people know about, understand and work on the key concepts, arguments and needs of EA.
Potential disagreement (shared as a total non-expert, to be clear)
My intuition is that AI safety research is still relatively undersupplied by social science researchers compared to the ideal. I think the area could, and ideally should, absorb a lot of social science research people over the next 30 years if funding and interest scale as I expect. Maybe 5,000+, if I consider all the organisations and geographies involved. Ideally, as many as possible of these people would be EA aware and aligned.
What might this look like? For instance, in academia, I see work to i) understand if/where creating evidence and interventions might be useful (e.g, interviewing/surveying technical researchers, policymakers and organisational leaders and/or mapping their key behaviours to influences), to ii) prioritise what is important (e.g., ranking malleability of different interventions/behaviour), then doing related research. I also foresee a range of theoretical work around how we can port over concepts and theory from areas such as communication, psychology and sociology to describe, understand and optimise how human and machines interact. I also expect that there will be a lot of value in coordination to support that work. I expect to see a global distribution of research labs researcher and projects.
In government and private settings, I foresee social science researchers hired to do more ‘context bound’ work with clear connections to immediate policy decisions. Government embedded research teams who need to understand technical, political and social factors to understand how to craft effective different types of national policy. Organisationally embedded teams engaged to create organisational policies that get value from internal AI, or engagement with other organisations platforms. Lots of work in the military and defence sector.
In support of these areas, I see a lot of social science background people being useful by working as knowledge brokers to effectively translate and communicate ideas between researcher and different types of practitioners (e.g., as marketers, community builders, user researchers, or educator). Also in providing various research support structures, training and curating potential students and research assistants, setting up support infrastructure like panels of potential technical/policy research participants, starting/organising conferences and journals etc).
I also suspect that’s there’s going to be lots needed that I have left out.
Overall, from a behavioural science research perspective (e.g., who needs to do what differently/what is the most important behaviour here/how do we ensure that behaviour happens), there are a lot of different behaviours, audiences contexts, and interactions, and little to no understanding of the key behaviours, actors, or ideal interventions. If this is life-on-earth-threatening-in-the-near-future stuff, then there is a lot of work to be done across a huge range of areas!
Not sure if this is an actual disagreement, so let me know. It’s useful for me to write up and share regardless, as it underpins some of my movement building plans. Feedback is welcome.