The race for AI to speak Catalan is heating up
In the space of a week, our language has gone from being an anomaly on Silicon Valley servers to becoming a standard function
BarcelonaWhen he opened WhatsApp on Thursday morning, he may have encountered a surprise that has gone largely unnoticed in the media, but which represents a minor earthquake in our digital lives. That iridescent blue circle that Meta embedded a few months ago at the top of chats has changed its behavior. Until four days ago, if he dared to address it in Catalan, Mark Zuckerberg's artificial intelligence (AI) would respond with a curt "no" in Spanish or Englishclaiming that it wasn't programmed for our language. It was a stubborn and technically absurd refusal, because the model that makes the device work (Llama 4) has been trained by reading the entire internet and, therefore, knows Catalan for some time now.
Well, without a press release, without official presentations with authorities in Barcelona, and without any announcement from the founder of Facebook, Meta AI has started speaking Catalan. And the most unsettling thing is not that it does it, but how it does it: with a frightening naturalness and a certain digital audacity. If you ask it about the change now, the chatbot has the option to tell you that it has always been there to help you, implying that perhaps before you "hadn't asked it properly." The machine wants you to believe that the problem was you and your lack of faith in the algorithm.
The elephant on WhatsApp
The importance of this move is enormous for a purely statistical reason: market penetration. It's often a mistake to think that everyone uses ChatGPT or Claude. The reality is that most people will never download a dedicated app to chat with AI, but instead, they live within WhatsApp.
Now the barrier to entry has disappeared. The aunt in Reus, the teenager in Girona, or the administrative manager in Lleida already have a Catalan-speaking assistant integrated into the tool they use a hundred times a day. Meta AI is no longer an external chatbot; it's just another contact in their address book, always available to write a congratulatory message, summarize the AFA group chat with 500 unread messages, or recommend a recipe.
However, we shouldn't get too carried away. The activation of Catalan in Meta AI hasn't been "formal." Despite what the chatbot promised, users haven't received any notification alerting them to the new feature. This is perhaps part of a low-profile strategy by Meta towards European regulators, who are scrutinizing how the company uses our data to train its models. This is the price of the party: privacy. Although personal chats remain end-to-end encrypted, everything you say directly to the AI is used to train it.
Furthermore, the system still makes mistakes. hallucinations —making up information with absolute confidence— are frequent. But the improvement will be exponential: the more we speak to it in Catalan, the more data we will generate for the system to refine itself and stop making spelling mistakes or inventing the history of Catalonia. Once again, we are de factoWorking for free to fine-tune a product for an American multinational. It's the price of comfort.
Salesforce: Companies speak our language too
While Meta AI is bursting into our private lives through WhatsApp, Salesforce has entered our lives as consumers this week with the formal addition of Catalan to Agentforce. The name might not ring a bell, but Salesforce is the invisible brain that manages customer service for most large companies, from banks to online clothing stores.
Until now, interacting with corporate chatbots was a guaranteed source of frustration. They were rigid systems ("If A calls you, answer B") that forced us to switch to Spanish right away. With Agentforce, the promise is different. They aren't just simple chatbots; they are "autonomous agents." This means they have the reasoning ability to understand the context and execute actions.
Let's imagine a practical e-commerce example. If you're in an online store and ask the assistant, "I'm looking for running shoes for Collserola that won't slip when it rains and that won't cost an arm and a leg," the agent should understand the geography (Collserola = mountain, dirt road), the weather (mud), the product (sunscreen with ad) and respond accordingly, instead of saying, "I don't understand 'Collserola.'" In the banking sector, these agents can handle issues like returned payments or blocked cards, all in standard Catalan and without having to wait for a human operator to be available at 3:00 AM.
The unjustified nostalgia for the human call center agent
This is where we must overcome a very ingrained cultural resistance: the urge to immediately ask to "speak to a human." Let's be honest: when you finally manage to speak to a human call center operator, that person is usually completely tied down, forced to read a solution manual on a screen. The current paradox is that the AI agent, connected to the entire database, often has more can real to solve your problem than the precarious human on the other end of the line. It's worth mentioning that Agentforce achieves this fluidity because, thanks to an alliance between Salesforce and OpenAI, it uses the ChatGPT engine behind the scenes, which works perfectly in Catalan.
Domino effect?
One week before the Salesforce announcement and the change in Meta, Google had officially announced support for Catalan —not yet fully implemented, it must be said— and 29 other languages in Gemini 3 Pro. The inclusion of Catalan in the same technical category as languages with their own status constituted a significant change in language policy on the part of Google.
The hypothesis that this decision by Google accelerated the unblocking of Catalan in Meta AI is plausible. When one of the tech giants incorporates a language, others face comparative pressure: maintaining language restrictions that the competition has already lifted generates criticism and damages its image. Meta, which had been resisting offering Catalan for months despite having the technical capacity, was facing an increasingly indefensible situation.
Why now?
Where does this sudden rush by digital giants to offer Catalan, Slovenian, or Hungarian come from? Let's not kid ourselves: they're not doing it out of love for Pompeu Fabra or cultural awareness. They're doing it because the very use of AI substantially reduces the cost of adding new languages, and besides, the market for dominant languages like English, Spanish, and Chinese is already well-served. The new race is to attract users of "non-dominant" languages with active online communities and a certain level of purchasing power. We are the "long tail" of the AI market, and they want us all in the same boat.
Is this normalization a victory? Undoubtedly: it improves the digital visibility of Catalan and prevents it from becoming a second-class language in the eyes of algorithms. But let's not forget the fine print: our linguistic normality now depends on switches located in California. If tomorrow Salesforce, Google, or Meta decide that maintaining the Catalan model isn't profitable, they can silence us again. For now, we'll take advantage of the conversation while it lasts.