Roughly twice as many published in 2025 as in 2024
75
Signals that labs, governments, and researchers want this work
Where social science fits
AI safety is broad, and social science does not cover all of it. Risks are commonly grouped into three families. We work on a deliberate slice of two of them.
Misalignment
Outside our scope
An AI system pursues something other than what its developers intended, as in reward hacking or deceptive behavior. This work needs deep machine learning expertise that the fellowship is not built to provide.
Misuse
Partly in scope
A person deliberately points a working system at harm, from cyberattacks to mass persuasion. We focus on persuasion and manipulation, where the harm runs through human psychology.
Structural risk
Largely in scope
Harm builds up from many actors’ choices with no single bad actor: concentrated power, labor disruption, eroding trust. We focus on its economic and institutional channels.
We also leave out state-level questions, such as AI races between countries and authoritarian surveillance. Within misuse and structural risk, we focus on four threat models where experiments, surveys, and field studies with real people give social scientists real leverage.
Why social science
Many of these harms run through people: what they believe, whom they trust, how they work, and how their institutions decide. Psychologists, economists, and sociologists already measure those things, using randomized experiments, large surveys, field studies, and administrative data.
That turns open questions about AI and society into findings that can be tested, replicated, and acted on by labs and policymakers. The examples below show what that looks like in practice.
A field that is taking off
Studies and reports in our database, by year of publication.
2020 through October 2026, with three earlier items not shown. Our search drew on lab and university publication pages, which favor recent work, so read this as a direction, not a precise count.
Four threat models, with real examples
For each threat model, two questions social scientists are already answering, and a study that answers them.
Psychology · Cognitive science
Human-AI interaction harms
Attachment to AI companions, over-reliance and automation bias, treating AI as human, and design that optimizes for engagement.
Does an AI companion make lonely people less lonely, or more isolated?
In a mixed-methods study of 404 regular companion-chatbot users, how much people used chatbots did not by itself predict loneliness. The researchers found seven distinct user profiles; companionship was linked to better well-being for some and worse for others.
Does a kind message still feel comforting once you learn an AI wrote it?
Across nine studies with 6,282 participants, identical empathic replies were rated as less caring and supportive once people learned an AI had written them. Given the choice, people preferred to wait for a human reply over getting an immediate AI one.
Psychology · Communication
Misuse: persuasion and manipulation
Disinformation, influence operations, and how susceptible individuals are to AI-generated persuasion. The same persuasive power that can correct false beliefs can also be used to plant them.
Can a short conversation with an AI talk someone out of a conspiracy belief, and does it last?
2,190 people who believed a conspiracy theory discussed it with GPT-4 Turbo, which offered personalized, evidence-based counterarguments. Belief fell by about 20%, and the effect was still there two months later.
Can AI tailor a persuasive message to someone’s personality, and does tailoring make it work better?
Across four studies with 1,788 participants, messages that ChatGPT tailored to people’s psychological profiles were significantly more persuasive than generic ones. Personalized persuasion, once costly, can now be automated.
Economics
Structural risk: economic and labor channels
Labor-market disruption, market concentration among AI firms, why companies cut safety corners under competition, and who captures value and control as AI spreads.
Is ChatGPT changing who gets paid more, or who loses their job?
Comparing occupations by how exposed they are to language models, before and after ChatGPT’s release, the authors found earnings rose in highly exposed occupations while unemployment did not change. So far, the adjustment has run through pay rather than job losses.
What do people using AI at work say it is doing to their jobs?
Open-ended responses from about 80,500 Claude users. About one in five raised concerns about losing work, and that concern rose with how exposed a person’s occupation is to AI; early-career workers worried most. Respondents chose to take part, so they are not representative of all workers.
Sociology · Management
Structural risk: institutional and organizational channels
Eroding trust in institutions, organizational risk, uneven AI adoption across groups and institutions, and the decline of shared knowledge as a social process.
Do AI hiring tools discriminate by age and gender, even when no one designs them to?
Across about 1.4 million online images and videos and nine language models, women were portrayed as younger than men, most of all in high-status jobs. When ChatGPT generated and rated about 40,000 resumes, it assumed women were younger and less experienced, and rated older men as more qualified.
What happens to professional judgment when a company gives its workers AI?
In a randomized experiment with 758 Boston Consulting Group consultants, those given GPT-4 completed 12.2% more tasks, 25.1% faster, at 40% higher quality on tasks within the AI’s abilities. On a task just beyond them, consultants using AI got the right answer less often than those without it: 60–70% against 84%.
Where the evidence is thin
Persuasion and manipulation has the least work so far: 36 of the 268 items, even though it is among the risks most directly measurable by experiment. Economics also dominates the database, with 99 items, against 54 from psychology and 41 from sociology. These gaps are where new researchers can add the most.
Beyond the research itself, AI labs, governments, and universities are building teams and funding programs for exactly this kind of work, and AI safety researchers have been asking for social scientists for years.
AI labs. Anthropic, Google DeepMind, Meta, Microsoft, and OpenAI each have dedicated teams, institutes, or programs studying AI’s economic and societal effects, and all five posted roles for economists and social-impact researchers in 2025–26.
Governments. The UK launched an AI Economics Institute in June 2026, and its AI Security Institute funds systemic safety research and runs a societal resilience program.
Universities. Stanford’s Digital Economy Lab, MIT Media Lab’s Advancing Humans with AI program, the Oxford Internet Institute, and the University of Virginia’s Economics of Transformative AI initiative all run dedicated research programs.
The AI safety community. Researchers have called for social scientists in AI safety since at least 2017, from Geoffrey Irving and Amanda Askell’s “AI Safety Needs Social Scientists” to recent posts on LessWrong and the EA Forum.
“This paper is a call for social scientists in AI safety.”
“I think we want to build a metropolis—a hub with dense connections to the broader communities of computer science, social science, and policymaking—rather than an isolated island.”
We searched the publication pages of AI labs, government institutes, academic centers, and individual researchers, and checked the authorship and relevance of each item before adding it. Each item is tagged by threat model and discipline. Studies with large samples or randomized designs are flagged as large-N or RCT. Press releases and short news items are left out of the research list.
This is a documented floor, not a census: there is certainly relevant work we have not found yet. Last updated October 2026. If you know of work we have missed, email contact@sira-project.org.
Fund or mentor this workHelp social scientists take on these questions.