TL;DR: I’m launching Societal Impacts Research on AI (SIRA), a mentored research fellowship to bring more social science talent into AI safety. We’re looking for our first cohort of 15-20 social scientists, from undergrads to postdocs, to work on societal impact problems within AI safety. The first cohort is projected to start in Summer 2027. Express interest as a fellow or mentor, or as a host organization or funder.
The Industrial Revolution spawned the birth of modern social science to explain how industrialization was reshaping work, institutions, and social life. Coming forth as a response meant that the frameworks social science aimed to answer operated after the fact, leading to a near-universal game of catch-up in making sense of the moment. Today, Artificial Intelligence is proving to be a disruption of a similar (or even larger) scale, and if this technology continues to arrive faster than we can explain and understand its impact, humanity can miss the opportunity to better steer its development and deployment. I believe that humanity needs researchers - specifically social scientists - to study AI’s effects on people and institutions while there is still time to shape the transition into a world of coexistence.[1]
Currently, social scientists who want to work on AI safety have no legible way in, except two weak options:
SIRA would be a 12-week, in-person research fellowship with a cohort of 15-20 social scientists.[3] At the conclusion of the program, fellows would leave with a finished project and a public presentation. Mentors would be selected first and come from an existing pool of researchers pursuing AI safety-relevant social science, from academia (faculty, postdocs, and PhD students) and from industry teams. Fellows will be matched in one of two ways: (1) Scoped projects: A mentor can offer a defined project to which fellows can apply directly to it. This path suits less experienced fellows who would benefit from a clear starting point. (2) Research directions: A mentor can offer a broader area to which fellows can then be matched by fit. This path suits PhD students and postdocs who have developed their own research agendas and have a moderate level of research experience.
Given that 12 weeks is a relatively short timeframe to conduct a social science research project, SIRA will likely prioritize people with more research experience. My current preference is toward postdocs, PhD, master’s, and undergraduate students with demonstrated research ability.[4] This take is grounded in the belief that in this situation, research experience probably matters more than AI safety context, as context can be upskilled in the timeframe leading up to the fellowship or learned early on.
To make SIRA maximize the experience for both fellow and mentor, I plan to borrow structural aspects from existing fellowships. For example, programs like MATS, Pivotal, and ERA have shown that certain structures can reliably turn mentorship into output, such as a dedicated research manager to help with project scoping, seminars and guest speakers to give fellows more safety context, in-person coworking to foster collaboration, financial support to allow fellows flexibility, and a closing symposium where fellows can showcase their research.
Ideally, SIRA would start as a stream inside an existing fellowship.[5] Doing this would allow SIRA to inherit certain structures like admissions, mentor networks, and credibility, and is the cheapest way to test the idea.
Despite the case for doing this now or even at all, SIRA is underpinned by several assumptions that the pilot will test. Some of them include:
From my perspective, there are three main reasons why this is the right time for a program like SIRA to exist. First, while some might assess the progress of AI development and view a social science fellowship as a purely long-horizon bet, I would argue that even if timelines are short, society has less time to understand and adapt to AI’s effects, and therefore should pursue research that seeks to fill this gap. Even Anthropic is making a similar wager by betting across multiple timelines so their efforts hold value in whichever future world comes to pass. Second, the field is already moving toward non-technical talent. Kairos recently argued that the biggest talent gap has shifted from technical researchers to generalists, and programs like MATS have even added non-technical tracks in response. In my view, the social sciences are a natural next group to invite to the table. Third, early-career talent is converting faster than expected. Kairos also reported that the most promising people in its programs are now reaching impactful roles in about 5-9 months, down from an earlier estimate of 1-3 years. If this holds for social scientists participating in SIRA, a summer 2027 cohort could be doing impactful AI safety work by 2028.
At present, SIRA exists as an idea in a doc (and now a LW post!). Here’s what you can do to help turn it into a reality:
Credit to Agus Covarrubias’s recent X post raising this topic through the lens of the Industrial Revolution.
In my view, PIBBSS is perhaps the closest existing program. It focuses on connecting fields like neuroscience, economics, game theory, philosophy, and physics to AI safety, largely to better understand AI systems.
SIRA’s focus is psychology, economics, and sociology. Adjacent social science fields such as communication (not broadcast/journalism), anthropology, cognitive science, and linguistics are welcome to apply. Further, by “social science,” I generally mean empirical research on humans and institutions, using methods like experiments, surveys, interviews, ethnography, and quantitative modeling, applied to AI systems and their effects.
This could be conference or workshop papers, or even preprints. Alternatively, it could be experience as an RA for a professor, or verified steps to pursue independent or mentored research.
The choice of where SIRA should be appended to rests on the existing program’s willingness to add another stream, take a risk on new research directions, and be the most optimal fit for safety context. While beggars can’t be choosers, I have a hunch that location will greatly inform the success of the fellowship, specifically that a location with a stronger safety culture and networks would likely produce the strongest research ideas and be the most beneficial to fellows gaining exposure. Some examples include: MATS (Berkeley, CA), Pivotal (London, UK), CBAI (Cambridge, MA), Collider (NYC, NY), or Kairos (potentially online or in Berkeley, CA).
This is probably the strongest objection, and one that is partly true. I do believe that a well-resourced, capable, and agile technical safety researcher could produce passable social science. However, I’m quick to counter that each discipline carries its own tacit knowledge and research taste. I’m a believer in letting each field do its own work, with cross-pollination, rather than asking one field to absorb another. Like Hamming’s suggestion of working with your door open, I believe that cross-pollination between social scientists and technical researchers is a field-level vision of working with the door open. This way, people from different disciplinary silos are exposed to each other’s methods and problems.
