Systèmes de modération de contenu par l'IA supprimant de manière disproportionnée le contenu francophone, autochtone et racialisé

AI-powered content moderation systems deployed by major social media platforms operating in Canada have repeatedly demonstrated disproportionate error rates when processing content in French, Indigenous languages, and content from racialized communities. Internal documents from Meta (revealed through whistleblower disclosures and reporting) showed that approximately 87% of the company’s spending to counter misinformation was devoted to English-speaking users, who represent roughly 9% of the platform’s user base. Non-English languages — including French — received substantially less investment in classifier training and human review capacity. This pattern extends across platforms: automated systems trained predominantly on English-language data systematically misclassify content in other languages, leading to both over-removal of legitimate speech and under-removal of harmful content.

Francophone Canadians — particularly in Quebec — use social media platforms where moderation systems may misinterpret Quebecois vernacular, colloquialisms, and cultural context. Indigenous language speakers face even starker gaps: content in Inuktitut, Cree, Anishinaabemowin, and other Indigenous languages is largely undetected by moderation systems that have little or no training data in these languages. The House of Commons Standing Committee on Canadian Heritage examined the conduct of major technology platforms in Canada, documenting how these companies resist regulation and how their algorithmic systems disadvantage Canadian content and linguistic minorities.

The pattern is ongoing rather than a single event. The Citizen Lab at the University of Toronto, in its submission on the federal government’s proposed approach to online harms, noted that people in Canada access content in hundreds of languages and dialects that do not receive equal moderation resources from platforms. Research consistently shows that platforms invest moderation resources roughly in proportion to advertising revenue rather than user population or rights impact, meaning languages and communities with less commercial value receive worse service. In the Canadian context, this raises questions about how the Official Languages Act’s guarantee of linguistic equality applies to digital platforms where an increasing share of civic discourse occurs.

Canada’s Online Harms Act (Bill C-63), introduced in February 2024, proposed a framework that could have begun to address these disparities through a Digital Safety Commission. However, Bill C-63 died on the Order Paper when Parliament was prorogued in January 2025, leaving Canada without dedicated legislation addressing platform content moderation obligations. As of early 2026, no Canadian legal framework requires platforms to demonstrate equitable moderation performance across languages or to report moderation accuracy disaggregated by language.