University of Toronto

Responsible & Critical AI

Midjourney images, prompted to analyze spatial conditions in an informal market in Dhaka. The generated images excluded women from the scenarios, which misrepre
Midjourney images prompted to depict an informal market in Dhaka: the generated scenes left out women, who are central to these markets. From In-Between Visuals and Visible (CHI 2024).

AI systems are increasingly used to inform, judge, and govern — often in ways that reproduce Western assumptions and existing inequalities. We evaluate how large language models behave, examine the politics of the data and methods behind AI, and study how AI is adopted and governed, especially in the Global South.

Our aim is AI that is accountable to the people it affects, grounded in non-ideal, real-world conditions rather than idealized ones.

Evaluating Large Language Models

Overview of our reasoning-based aggregation pipeline.
Connecting Causes & Failures & Mitigations. Sankey diagram connecting contributing causes to associated failures and mitigations.
Left: a reasoning-based aggregation pipeline for multi-perspective ethical judgments (Canadian AI 2026). Right: how contributing causes connect to LLM failures and mitigations in software engineering (TOSEM 2026).

As LLMs are increasingly used to judge bias, they may themselves show bias in subtler ways. Evaluating Second-Order Bias of LLMs Through Epistemic Entitlement (EMNLP 2026 Findings) introduces a philosophically grounded way to measure social bias in an LLM's judgments about social bias. We have also tested how LLM-backed chatbots handle misinformed and demographically informed prompts about climate change and mental health (AI Magazine 2025).

We study how people experience these systems in practice: how users flag and respond to LLM sycophancy, a fair, multi-perspective ethical reasoning framework, and why software engineers sometimes find LLMs counterproductive (TOSEM 2026).

Related papers

Data, Annotation, and Ground Truth

An imagination of how our methodological framework can be presented. To illustrate, consider “Feature dependencies” as an undesirable property mapped to the rea
(a) Distribution of survey participants’ primary and secondary (if selected) research areas.
Left: a non-ideal methodological framework for responsible ML (CHI 2024). Right: survey participants' research areas (AIES 2026).

"Ground truth" assumes there is one correct label. The Consensus Trap (FAccT 2026) reviews five years of research to argue that this paradigm treats human disagreement as noise rather than a vital sociotechnical signal. Annotation is also political: Data, Annotation, and Meaning-Making (FAccT 2024) shows how annotation overlooks faith-based, religious, and spiritual aspects, drawing on a dataset of faith-based communal violence.

We also look at how AI work gets done. Towards a Non-Ideal Methodological Framework for Responsible ML (CHI 2024) draws on interviews with 22 ML practitioners; Studying People to Study AI (AIES 2026) surveys 93 experts on the place of human research in AI safety and ethics; and UnWEIRDing Peer Review (CHI 2026) examines how review biases disadvantage Global South scholarship.

Related papers

AI, Power, and Governance

Top: Distribution of AI systems across GC organizations; Bottom-left: Provinces and Territories explicitly mentioned in the AI register using population-normali
Low-literate local graphic designers with their unlimited imagination yet limited vocabulary struggle to produce text-based prompts. Their ideas get marginalize
Left: AI systems across Government of Canada organizations and the provinces named in the Federal AI Register (FAccT 2026). Right: local graphic designers working with text-to-image tools (CHI 2024).

Transparency tools are not neutral. Bureaucratic Silences (FAccT 2026) analyzes all 409 systems in Canada's first Federal AI Register to show what such registers reveal, omit, and obscure. Prediction Laundering (AIES 2026) examines how prediction markets such as Polymarket produce seemingly objective "truth" that hides uncertainty, manipulation, and capital asymmetries.

In the Global South, we study Bangladesh's AI readiness across curriculum, infrastructure, and governance (COMPASS 2026), the ethical risks of computer vision applications (GI 2020), and how text-to-image tools reshape creative work — from the impacts on Bangladeshi image-makers (CHI 2024) to community-centric design of T2I tools (AIES 2025).

Related papers

Talks

Ishtiaque Ahmed on whose values AI systems encode, and how AI ethics changes across cultures and borders: "Whose intelligence? Whose ethics?" at the Schwartz Reisman Institute seminar series, and "Ethics in Translation: AI Across Borders" at the Centre for Ethics, University of Toronto.