University of Toronto

Online Harms & Moderation

Feedback from Mod-Guide’s in Mediator role.
Mod-Guide, an LLM-based feedback system that helps moderate insensitive speech toward Indigenous ethnic and religious minority communities, shown here in its Mediator role. From Mod-Guide (COMPASS 2026).

Online platforms are where many people in the Global South meet, organize, and get their news — and where they face harassment, hate, and misinformation. Mainstream moderation is built largely around Western, secular values and well-resourced languages.

We study how harms are experienced by women, religious and ethnic minorities, and other marginalized users, build detection resources for low-resource languages, and design moderation that draws on community values.

Harassment, Gender, and Justice

Women in the Global South often seek justice for online harassment by sharing screenshots as evidence — evidence that is frequently challenged for its authenticity. Based on a survey of 91 and interviews with 43 Bangladeshi victims, we built Unmochon (CHI 2021, Best Paper Honorable Mention), a tool for collecting trustworthy evidence of harassment on Facebook Messenger. Earlier, we studied harassment on anonymous social media in Bangladesh (ICTD 2019).

We have examined who participates in movements like #MeToo and why many do not, in both Bangladesh (GROUP 2020) and India (CSCW 2021). In a conservative society where talking about sex is taboo, ShishuShurokkha (CHI 2022) takes a transformative-justice approach to combating child sexual abuse.

Related papers

Hate Speech, Misinformation, and Polarization

Heatmap demonstrating the relationship between type of hate and severity .
Source of Getting News
Left: how the type of hate relates to its severity in Bangla social media (ACL 2026). Right: where people in Bangladesh get their news (CSCW 2020).

Reliable detection is especially hard in low-resource languages. LLM-Based Multi-Task Bangla Hate Speech Detection (ACL 2026) detects the type, severity, and target of hate speech in Bangla, and POLAR (ACL 2026 Findings) is a benchmark of over 110K instances in 22 languages for studying online polarization across cultures and events.

We also study the people behind the data. Combating Misinformation in Bangladesh (CSCW 2020) examines the roles of journalists, fact-checkers, and users; The Politics of Fear (CSCW 2024) documents the experiences of religious minority communities on social media; and Combating Islamophobia (TSC 2023) designs for compromise, community, and harmony in mitigating harmful content.

Related papers

Community-Centred Moderation

Prompt, LLM, and RAG pipeline.
Dimensions and stratification of usership
Left: Mod-Guide's prompt, LLM, and retrieval pipeline (COMPASS 2026). Right: dimensions of usership for intersectionally marginalized users (ICTD 2024).

As moderation increasingly relies on large language models, we ask whether these systems can recognize culturally insensitive speech toward historically underrepresented communities. Mod-Guide (COMPASS 2026) is an LLM-based moderation feedback system built for this purpose, and Argument-Based Consistency in Toxicity Explanations (EACL 2026 Findings) evaluates how LLMs reason about toxicity, not just whether they detect it.

Community values matter too. A civics-oriented study (ICTD 2024) shows how intersectional identities shape Indigenous and women users' experiences of hate speech; a three-year rural ethnography, Toleration Factors (ICTD 2022), examines expectations of decorum, civility, and certainty; and Imagined Online Communities (CSCW 2022) studies how Facebook group admins govern their communities.

Related papers