While major AI chatbots can generate text in Catalan, experts are sounding the alarm that they lack genuine cultural understanding, often translating from an English-centric worldview rather than “thinking” in Catalan. To address this, a new initiative plans to have these AI models sit official language and culture exams, including the university entrance test, to measure and improve their cultural fluency.

Your browser does not support the video tag.

Accent Obert, a civil organisation dedicated to promoting the Catalan language in the digital world, spearheads this plan. The group, formerly known as Fundació .Cat, has identified a critical flaw in how current generative AI models operate. While the twelve leading commercial chatbots can understand prompts in Catalan, the process behind their replies is often a multi-step translation.

“The joy is that AIs respond in Catalan, but the problem is that they don’t think in Catalan. They think in Wisconsin and translate to Catalan. We need responses built from the Catalan cultural framework,” explained Genís Roca, president of Accent Obert, in a recent presentation.

This translation process-where a Catalan query goes into English, gets processed, and then re-translated back for the user-can strip away nuance and embed foreign cultural biases. As Roca highlights:

“What does Joan Miró represent for Catalan culture? I don’t want an American to answer that for me.”

A Plan for Cultural Fluency

To tackle this challenge, Accent Obert has unveiled an urgent plan. By 2026, the organisation intends to subject major AI models to the same official exams that Catalan citizens take. These will include basic competency tests, the rigorous university entrance exams (selectivitat), and advanced language proficiency tests up to the C2 level.

Accent Obert will use the results to create a public ranking of AI models based on their understanding of Catalan language and culture. This aims to objectify their capabilities and incentivise developers to create more culturally attuned systems. According to technologist and Accent Obert ambassador Albert Cuesta, while twelve chatbots can understand Catalan, only three can currently hold a sustained conversation in the language. He is confident this number will grow.

A core part of the strategy involves creating a “federation of Catalan content.” This curated library will serve as a definitive resource for an AI when a user asks it to respond from a Catalan perspective, effectively teaching it “what it means to be Catalan.”

Building on a Strong Digital Foundation

The initiative builds on decades of successful digital activism in Catalonia. Roca highlights the region’s robust online presence as a key advantage.

“For 20 years we have had the .cat domain, the first cultural domain in the world,” Roca noted, pointing to its record 116,000 active domains.

The Catalan Wikipedia is also a 25-year-old institution, one of the first language versions for a stateless culture.

Crucially, projects like Projecte AINA have significantly advanced the technical groundwork. Managed by the Barcelona Supercomputing centre and backed by the Government of Catalonia, AINA has multiplied the Catalan language data available for training AI models by a factor of 100. This is complemented by the long-standing work of Softcatalà, a non-profit that provides essential tools and software for using Catalan in technology.

This history of proactive engagement gives advocates hope.

“We did a study, told Google they had a problem they didn’t know they had, and they fixed it for us in twenty-four hours,” Cuesta recalled, noting the fix also helped 130 other languages.

Addressing AI-Specific Challenges

Despite this optimism, Accent Obert acknowledges three key risks in the AI era. The first is a dependency on models developed outside the Catalan-speaking world. The second is the cultural impact of AI systems that operate with alien cultural frameworks. The third is the risk of “linguistic friction,” where users might revert to English if they find it yields more accurate or comprehensive AI responses.

“The internet has a tendency to flatten cultures. Cultures must defend themselves,” Roca stated.

The belief is that tech companies will be receptive, as customisation is a key driver in the industry.

“The digital industry kills to personalise the user experience, and nothing personalises more than treating people in their own language,” he added.

The initiative is one of several fronts for Accent Obert, which also works with the City Council of Barcelona and promotes digital content creators through its Crit awards. As reported by VilaWeb and Ara, the ultimate goal is ambitious: to ensure the Catalan language not only survives but thrives with authenticity in the age of artificial intelligence.