Entraînement de l'IA sur des œuvres protégées et l'économie créative du Canada

Frontier AI systems are trained on vast corpora that include copyrighted works by Canadian creators — books, journalism, music, visual art, code, and academic research — without consent, compensation, or attribution. This creates a dual hazard: it may violate creators' existing rights under Canadian copyright law, and the resulting AI systems increasingly substitute for the creative labour that produced the training data, eroding the economic foundations of Canada's cultural industries.

The legal landscape is actively contested — and Canadian litigation is now at the forefront. In November 2024, a coalition of Canada's leading news publishers — The Canadian Press, Torstar, The Globe and Mail, Postmedia, and CBC/Radio-Canada — sued OpenAI for copyright infringement through unauthorized scraping of news content to train ChatGPT. In November 2025, the Ontario Superior Court ruled it had jurisdiction to hear the case, rejecting OpenAI's motion to dismiss and awarding the plaintiffs $260,000 in costs. This is the first Canadian case to directly address copyrighted content use for AI training.

Separately, Vancouver author J.B. MacKinnon filed four class actions in BC Supreme Court in 2025 against Meta, Anthropic, Databricks, and NVIDIA, alleging they used "the Pile" — an 800-gigabyte dataset containing roughly 196,640 unlicensed books, many by Canadian authors — to train their models. In total, no fewer than six class actions were proposed across Federal and Provincial Courts in Quebec and British Columbia in 2025. Internationally, Anthropic agreed to a $1.5 billion settlement in Bartz v. Anthropic (August 2025), covering approximately 500,000 pirated works — the largest copyright settlement in U.S. history — with direct implications for Canadian authors whose works were in the training datasets.

Canada's Copyright Act does not contain a text-and-data-mining (TDM) exception for AI training, unlike Japan, the UK, Singapore, or the EU's conditional exception. The government launched a formal consultation on copyright and AI in October 2023, receiving close to 1,000 responses. ISED published its "What We Heard" report on February 11, 2025, finding no stakeholder consensus: creators argued unauthorized AI training violates existing law, while technology companies contended TDM extracts factual patterns rather than expressive content. No Copyright Act amendments have been tabled as of March 2026.

The Standing Committee on Canadian Heritage (CHPC) commenced a study on the effects of AI on creative industries in September 2025, hearing from the Coalition for the Diversity of Cultural Expressions, Copibec, SOCAN, and others. The committee moved to drafting instructions in November 2025. Canadian creators' organizations are unified in demanding a licensing regime. SOCAN stated that AI companies "ingesting and training models on copyrighted musical works without permission from, compensation for, or credit to creators is not fair use, but theft." The Writers' Union of Canada stated unequivocally that "AI training, including text and data mining, on published material without author permission is copyright infringement." Access Copyright, Copibec, and Music Publishers Canada have taken similar positions.

Canada's creative industries contribute approximately $55.5 billion annually to GDP and employ over 600,000 people (2020 figures). The structural risk is that AI systems trained on the output of these industries erode their economic viability, creating a paradox where the training data pipeline depends on an industry that AI is simultaneously undermining.