Quebec’s language laws face a new reality online: automatic translation

Quebec’s new premier Christine Fréchette meant to post on X about a very Montreal kind of evening.

Her first appearance as premier on Tout le monde en parle, Radio-Canada’s flagship talk show? Check.

The Canadiens’ 4-3 win? Check.

“Une très belle soirée!” (a great evening) she concluded.

But for English-reading users relying on the platform’s automatic translation, the post landed differently, if a bit arrogantly.

X’s translation, powered by its AI chatbot Grok, took the title of the show at its literal English translation, which means, Everybody’s talking about it. It rendered the post as: “Everyone’s talking about her as Prime Minister.”

Automatic translation is now built into X by default, meaning users receive content in the language they prefer, as adjusted in their settings. The platform’s head of product, Nikita Bier, said in April the company was rolling out auto-translate worldwide “to give posts in any language global reach.”

Other tech companies have also been moving in a similar direction, leaning into artificial intelligence to provide translations by default. Apple’s latest software prompts users to automatically translate messages. TikTok, too, has been rolling out automatic machine translation across posts.

In Quebec, French by law is required in workplaces, commerce, public signage, and must be no harder to access than any other language. The responsibility rests on those who post it.

So, what happens when language online is available but sometimes inaccurate? What risks do businesses face in relying on platforms to translate? And what does that mean for connection across communities that speak different languages?

OQLF opens door to auto-translation

Quebec’s language watchdog, the Office québécois de la langue française, told The Gazette that commercial social media posts by companies about products or services intended for the Quebec market must be available in French, and that machine translation can be one way of doing that.

“The OQLF believes that using innovative means based on information and communication technologies can help ensure a greater presence of French,” it said in a statement. “If the automatic translation module allows users to access commercial publications in French, without any changes to the settings, regardless of the device used to view them, then these publications would be considered available in French.”

The office added that compliance must be assessed case by case.

“To assess the compliance of commercial publications on a company’s website, for example, a text is considered not to be in French if, to understand it, one must refer to its version in another language.”

Asked whether the OQLF has analyzed the quality of translations, it said: “The OQLF’s mandate does not include analyzing the language quality of content generated by automatic translation tools.”

A review by The Gazette of X posts translated between both English and French found most translations were accurate. Some, however, did not provide the intended meaning. Others were not translated at all.

The OQLF’s current framework was not designed with automatic translation in mind, said Julianne Chu, a lawyer and translator for Éducaloi, and it also does not distinguish between how French is produced, whether it has been human translated, machine translated or even done automatically.

“It really focuses on whether the French translation is available and meets legal requirements, notably in terms of quality and accessibility.”

For businesses, she said, it presents a confusing picture and carries risks, particularly when relying on a platform to handle translation.

How do machines translate language?

Grok’s translations are powered by what are known as large language models, the same kinds of systems behind tools like ChatGPT, Google’s Gemini and Anthropic’s Claude. Most operate as what researchers call “black boxes”: they generate an output, but offer little insight into how it was produced. They are trained on vast amounts of text to learn how language works. In 2025, nearly six in 10 Canadians have used an AI tool like Grok or ChatGPT, according to a Leger survey.

For translation specifically, that training typically happens in two stages, according to Jackie Cheung, a computer science professor at McGill University who researches how AI understands and generates human language.

First, he said, the systems are exposed to enormous volumes of text in English, French and other languages, allowing them to see how words tend to appear together, how sentences are structured, and how meaning is usually conveyed.

Then comes a second stage, more specific to translation. The models are trained on pairs of text, and learn how to map one onto the other.

At a basic level, the systems are learning patterns. However, whether these systems truly understand what they are translating, he said, remains a “thorny” question in the field of AI research.

Still, they can be effective at translating the surface of language, he said, especially in more standard or formal contexts.

“What’s kind of missing is the context necessary to understand the intentions of the original poster fully,” Cheung said.

And that is particularly true online. Language often depends on shared references — memes, in-jokes and tone — that exist within a community. Outside that context, a translation can feel off or misleading, even if it is technically accurate.

Montreal, he noted, has its own version of that, too, where sometimes it may not even be pure French or English spoken or written but a mix of both.

“The choice of language that you use itself contains information,” Cheung said. “And this would, by definition, be lost if you have auto-translate on.”

The result is a risk that regional expressions, dialects or ways of speaking are smoothed into something more standardized, more aligned to how the generalized models were trained on.

Why AI models may overlook Quebec

Eeham Khan, a PhD researcher at Concordia University, is currently building a Quebec-first language model to counter that.

“It’s not always just about language,” he said in an interview with The Gazette. “The data that we have is not just the language of Quebec, but it also represents the terms, the cultures, things that maybe other models or larger models might not know about too well.”

When he began working on the project last year, he recalls asking ChatGPT to translate something into Quebec French. The chat bot’s reply was like “it’s forcing itself to be the most ‘redneck Quebecer’ imaginable.”

“It’ll use eight different slang terms in two sentences,” he said. “You read it and think: no one talks like this. No one writes like this.”

“It knows what the idea of Quebec is, but it takes it to the extreme.”

The biggest problem, he said, is data availability.

Khan is trying to build that foundation, working with partners including Radio-Canada to gather data with consent.

“Even with all of that, we can’t really hope to compare to the big models,” he said, pointing to the time and resources required to collect data ethically. And as those larger systems become more widely used, he said, they begin to shape how language is expressed.

According to Cheung, Indigenous languages raise even more sensitive questions. Some communities may not want their languages fed into translation systems, especially if the result is more access for outsiders than benefit for the community itself.

“The other issue is that it’s not clear that the same pipeline that works for English and French actually works for other languages where there’s not so much data available,” he said, adding that the copious amounts of data that the models heavily rely on may not exist for many other languages.”

In April, draft regulations tabled by Marc Miller, Minister of Canadian Identity and Culture and Minister responsible for Official Languages, set out how federally regulated businesses must provide services in French, targeting sectors that have historically slipped through the cracks of Quebec’s Charter of the French Language because they fall under federal jurisdiction.

The Department of Canadian Heritage told The Gazette: “Language compliance requirements remain the same, but the means to fulfill them are evolving. This was taken into account in the analysis leading to the draft regulations that pertain to the private sector.”

It added: “The obligations for federally regulated private businesses that would fall under the jurisdiction of the draft regulations would be to provide communications and services to consumers in French, regardless of the means used. The use of French would need to be at least equivalent to the use of any other language.”

Governments elsewhere in the world are already starting to respond to X’s move. Just in April, the U.K. Embassy in Japan posted that it was not responsible for Japanese translations automatically generated on X.

Ultimately, Cheung says he remains cautious about its potential to bridge connection.

“You can even see it within English,” Cheung said. “There are so many sub-communities, and there are still massive misunderstandings and polarization.”

“I don’t think it will help that much in terms of increasing connection or inter-community understanding,” he said. “Language is an important part of it. But it’s not the whole thing.”

As for Khan, he says he “understands the benefits of having like these translation services,” and that “they’re good in promoting accessibility.”

“But on the other hand, this only works really if the translation service works correctly. If you’re mis-translating information from politicians or CEOs coming from these like smaller countries, smaller provinces, etc, who are tweeting in their native languages or native dialects, it could potentially be very damaging for those people or for those cultures.

“So, like everything, it has to be done, right. It has to be done responsibly. It has to be done carefully.”

Do you have story tip? Write to me at hnorth@postmedia.com

The post Quebec’s language laws face a new reality online: automatic translation appeared first on Montreal Gazette.