Inconsistently. AI tools often default to international French rather than Québécois usage, and no Canadian source publishes an accuracy rate for it.
Short answer
Inconsistently. Mainstream AI tools have improved a lot at French generally, but what they produce often reads as international or European French rather than the Québécois or federally-styled Canadian French your reader expects — an anglicism here, “un email” where Quebec convention favours “un courriel” there. No Canadian regulator publishes an accuracy rate for this, so treat fluent-sounding French output as a first draft, not proof it's correct for a Quebec audience.
It comes down to what a language model is trained on. Canada’s Cyber Centre makes the underlying point in its own generative-AI guidance: most training datasets are pulled from the open internet, and “generated content has a fundamental bias in that only limited amounts of the world’s total data are online and available for AI to use.” (Canadian Centre for Cyber Security, ITSAP.00.041) Whatever variety of a language is most common online ends up over-represented in training, and international and European French — a far larger share of the world's online French — crowds out a smaller regional variety like Canadian French.
The tells are usually vocabulary, not grammar: an anglicism carried straight across from English, or a France-French word choice where Quebec usage differs. This is a language-quality question, not a licensing one, and it's worth being honest that no Canadian body has published a figure for how often it happens with any given tool — a claimed accuracy percentage here would be invented, not sourced.
Quebec's own language authority, the Office québécois de la langue française, maintains reference tools built specifically for correct Quebec-French terminology. If a piece of AI-drafted French is going in front of Quebec customers — a product page, a contract, a notice — check it against a Canadian French reference the way you'd check any translation, rather than trusting how fluent it sounds. See also why a confident-sounding answer isn't the same as a checked one and the same training-data gap showing up as a different kind of error.
See how brand voice and quality control actually get checked before customer-facing copy ships.
