Social theory and agent design
Socially grounded architectures and coordination
Investigating how social roles, institutions, and plural perspectives can inform the design and coordination of agentic systems.
University of Toronto · Social Agentics Lab
Agentic AI.
Socially situated.
The Social Agentics Lab investigates how agentic AI systems are shaped by, and participate in, social and organizational life. Bringing social theory into dialogue with technical design, our research examines social cognitive architectures, multi-agent simulation, and the implications of AI for power, equity, and accountability.
Research areasUnderstanding systems through their relationships
01 — The lab
Social theory and
agentic system design.
The Social Agentics Lab brings together researchers in information studies, computer science, and the social sciences to study the design and use of socially situated agentic AI.
We explore how social knowledge becomes part of agent design, and how agents can be situated within existing cultural and organizational systems. Our work connects social theory with the development of agent infrastructures and cognitive architectures.
02 — Research areas
Researching socially situated AI.
We examine how social knowledge, cultural differences, and organizational practices shape the design, behaviour, and evaluation of agentic systems.
Cultural representation and evaluation
Population fidelity and cultural differences
Evaluating how language models represent diverse populations, preserve differences between groups, and respond to culturally situated perspectives.
AI, knowledge, and social organization
Humanness, labour, and accountability
Examining the assumptions about people and knowledge embedded in AI, the human labour sustaining automation, and their implications for power and accountability.
Current grants
Funded projects supporting the lab’s research.
SSHRC · Insight Development Grant · 2025
Academic knowledge and generative AI reference services
Research on academic knowledge and generative AI reference services in libraries. Two-year project. Applicant: Matt Ratto; co-applicant: Kate Davis; collaborators: Sabina Pagotto and Catherine Steeves.
Project overviewCIFAR · CAISI Catalyst Project · 2026
Towards Socially Grounded AI Safety: Integrating Causal and Institutional Reasoning in Language Models
Developing agentic architectures and causal reasoning modules that incorporate cultural norms and institutional context, with an emphasis on relational accountability and trustworthy human–AI coordination. Collaborators: Matt Ratto and Zhijing Jin (Canada CIFAR AI Chair, Vector Institute; University of Toronto).
Project overview03 — Publications & presentations
Publications and presentations.
Papers, talks, and workshops on situating agentic AI within human social worlds.
Publications
2026 · arXiv preprint
Population Fidelity: Evaluating Population Representativeness in LLMs
A framework for evaluating whether LLMs reproduce group-level survey responses and the magnitude and structure of differences across populations.
2026 · arXiv preprint
Socially Grounded Agentic AI: Coordinating Plural Perspectives through Social Theory
A design perspective on pluralistic alignment: coordinating roles, perspectives, and feedback through social theory.
2026 · ICML Pluralistic Alignment Workshop · Non-archival paper
The Wedge Questions: Latent Cultural Boundaries in LLMs via Persona Projection Divergence
Examines cultural distinctions in model representations through Persona Projection Divergence, including differences that similar textual responses can obscure.
2025 · CSCW Companion
Social Agentics
A research agenda for situating agentic systems within specific social and organizational contexts.
2025 · ACM COMPASS
Social Agentics: ACM COMPASS workshop
A workshop contribution bringing social and organizational perspectives into the design of agentic AI.
2024 · GenAICHI workshop at CHI · Position paper
Posthumanist AI: Rethinking ‘the human’ as a model for generative AI
Investigates whose traits and behaviours become models for generative AI, and how broader accounts of humanness can inform its design and evaluation.
2022 · CHI Extended Abstracts · Workshop paper
Outsourcing Artificial Intelligence: Responding to the Reassertion of the Human Element into Automation
Examines the human labour sustaining AI and automation, and the implications for work, responsibility, and system design.
Presentations & workshops
September 19, 2025 · iSchool Institute · Recorded talk
Social Agentics: the challenges and opportunities of agentic AI
The possibilities and challenges of agentic AI, and why social and organizational context matters for its design.
2025 · ACM CUI · Keynote
Conversations in Context: Considering the Sociality of Agentic AI
How power, social structures, and norms can inform agents capable of navigating dynamic human worlds.
July 23, 2025 · ACM COMPASS · Workshop
Social Agentics: Situating Agentic AI Within Social & Organizational Contexts
An interdisciplinary workshop exploring social knowledge, agent architectures, and responsible approaches to agent design.
Writing & Commentary
Recent writing
Research reflections, commentary, and updates from the lab.
· Lab announcement
Introducing the Social Agentics Lab
The lab’s founding announcement outlines a research agenda connecting social cognition, multi-agent simulation, and the social and organizational implications of agentic AI.
Read on LinkedIn04 — Tools
Make the system visible.
Resources for agent design, population representativeness, and cultural evaluation.
Design toolkit · Paper & digital
AgentGraphs
Map an agent as connected, bounded loops. Explore the models, tools, memory, and controls each loop needs, and trace how work passes between them.
Public repository · Evaluation
Population Fidelity
Python and R pipelines for evaluating how faithfully LLM responses represent human populations. Compare survey responses across groups, measure the magnitude and structure of between-group differences, and generate analysis tables and figures.
Read the Population Fidelity preprint ↗Public repository · Fine-tuning
Culture Steering via Fine-tuning
The companion training pipeline for population fidelity research. Uses World Values Survey responses to create cultural and distributional LoRA fine-tuning experiments, providing adapters for comparisons in the evaluation framework.
05 — People
A shared inquiry.
Matt Ratto
Director · Professor, Faculty of Information
University of Toronto
Faculty affiliates
Profiles coming soon.
Students
Profiles coming soon.
06 — Student opportunities
Student research positions
Research jobs for students will be posted here as they become available, with project descriptions, eligibility criteria, and application details.
There are no advertised student research positions listed at this time. Please check back for future postings.
07 — Connect
Research starts
with a conversation.
We welcome conversations with organizations and researchers interested in shaping socially situated agentic AI.
Get in touch