What happens when algorithms enter public institutions?
We pair technical, computational methods with critical, interpretive inquiry, and we ground every project in a long-term partnership with the people who live with the system. The four areas below hold most of our current work. We also study housing and homelessness, higher education, algorithmic policing, and online safety in the Global South, and the publications page has the full list.
Statistics and machine learning: generalized linear models, predictive modeling, causal inference, and audits of systems already in deployment.
Ethnography, interviews, document and content analysis, and participatory co-design with workers, administrators, and affected communities.
The allocation of risk in child welfare
Our longest-running work embeds inside child-welfare agencies to study how caseworkers and predictive risk models interact, where automated risk scores distort street-level decisions, and how to design accountable alternatives with the workers and families involved. It has been cited in the UNDP Human Development Report and the ACLU's audit of the Allegheny Family Screening Tool, and continues with the Children's Aid Society of Toronto.
The accountability gap in public-sector AI
We study how governments adopt, register, and answer for the algorithms they deploy, from Canada's Algorithmic Impact Assessment and public AI register to the large language models now reaching frontline public services. This work is the basis of the forthcoming book Public Interest Technology: When AI Becomes Government.
The design of participatory health AI
With Peel Public Health, CIFAR's AI for Diabetes Prediction & Prevention network, and a CIHR team, we build and audit predictive tools for clinical and population health, and study how to involve the communities they affect rather than reduce them to a single risk score. Recent projects include patient-experience NLP and fairness in diabetes risk models.
The measurement of bias in language models
We audit how large language models encode identity, bias, and toxicity for communities they were never designed to serve, with sustained work across transnational Bengali communities and open datasets that make non-Western NLP bias measurable. Recent projects examine political bias in GPT-4 and whether LLM toxicity explanations hold up.