Latin America Builds Its Own AI To Rival ChatGPT
The Problem When Global AI Fails Locally
One morning in Santiago, Chile, graphic designer Juan Palma turned to ChatGPT for a simple task: directions to a nearby subway station. The response he received was swift, confident, and completely wrong—it sent him in the opposite direction. Palma’s experience highlighted a growing realization across the continent: the world's most popular AI models are often out of their element in Latin America. "I was surprised by the orientation error in ChatGPT’s response," he noted. "I had different expectations for the accuracy of this tool."
This isn't an isolated incident. While large language models (LLMs) like GPT and Meta's Llama are trained on multilingual data, they struggle with regional dialects, local idioms, and cultural context. This can lead to everything from simple inaccuracies to more significant fabrications, often called hallucinations.
A Solution Born in Latin America Introducing Latam-GPT
Frustrated by these shortcomings, a coalition of over 30 institutions across Latin America has embarked on an ambitious project: Latam-GPT. For the last two years, this Chile-led initiative has been developing an open-source LLM designed to serve the region's unique needs. Scheduled for a public release in September, Latam-GPT is being trained by locals who understand the linguistic and cultural nuances that global models miss.
Héctor Bravo, an expert in disruptive technologies, described the project as "building AI in Latin America, for Latin Americans." He emphasized that success isn't just about speed or accuracy, but about "cultural representation, social impact, and accessibility."
The Latam-GPT project is building AI in Latin America, for Latin Americans.
A key feature of Latam-GPT is its commitment to deep multilingualism. Beyond standard Spanish and Portuguese, it is being trained on Indigenous languages like Nahuatl, Quechua, and Mapudungun, as well as various regional dialects, including those from the Caribbean.
Following a Global Trend of Regional AI
Latin America is not alone in its quest for a more localized AI. This movement is part of a broader global trend. In Southeast Asia, the Sea-Lion family of LLMs is trained on nearly a dozen regional languages. In Africa, UlizaLlama interacts with users in at least five different languages, including Xhosa and Zulu. Meanwhile, India's BharatGPT supports over 14 regional languages, and the Indian government has announced plans to build its own sovereign LLM.
Laying the Groundwork for Regional AI
While Latin America has been relatively slow to adopt AI, it's rapidly catching up. A 2025 UN report identifies Chile as a leader in AI regulation and development. It's no surprise, then, that the idea for Latam-GPT emerged from Chile’s National Center for Artificial Intelligence (CENIA), founded in 2021.
"Given the scale of the initiative, we knew it was something that required broad collaboration," said Alvaro Soto, head of CENIA. His team fostered an "open spirit," recruiting universities, government offices, and civil organizations. Last year, CENIA secured 33 strategic alliances across Latin America, the U.S., and Spain, assembling a model with 50 billion parameters—comparable to ChatGPT 3.5.
The Latam-GPT Advantage Context is Key
What sets Latam-GPT apart? While global LLMs support Spanish, their training data often comes from Spain or is translated from English, missing crucial local context. In contrast, Latam-GPT is being trained on data sourced directly from the region's schools, businesses, libraries, and historical texts. This approach "helps the model better understand the contexts and needs of Latin American users," explained Omar Florez, the project's technical lead.
This is critical as generative AI use soars in the region. Brazil is already the third-largest user base for ChatGPT, and Meta's Llama downloads are surging. From classrooms to customer service and even drafting legal rulings in Buenos Aires, the demand is undeniable.
Overcoming Hurdles on the Road Ahead
The project is not without significant challenges. Latam-GPT's resources are dwarfed by those of its global competitors, and for now, it will be a text-only model. "It requires ultra-high-capacity infrastructure, specialized talent, and relevant data sets—three areas where gaps still exist," noted Carlos Honorato, CEO of an AI company.
Environmental concerns also loom large. LLMs are notorious for their high energy and water consumption. The infrastructure for Latam-GPT is located in a drought-stricken region of northern Chile, raising local concerns about data center impacts. However, the CENIA team states they are mitigating this by using a scalable cloud infrastructure powered by solar energy, which should significantly limit its environmental footprint.
Additionally, inconsistent data privacy laws across Latin America present potential legal risks. Experts warn that mishandling personal information could lead to litigation and reputational damage.
To be successful, Latam-GPT will need to ensure the participation of Indigenous peoples, migrant communities, and other historically marginalized groups.
The Ultimate Goal Representation and Regional Pride
Perhaps the most critical challenge is ensuring true representation. While Latam-GPT is a step forward, experts stress the need to actively involve Indigenous peoples, migrant communities, and other marginalized groups in the model's development and validation to avoid perpetuating biases.
Rodrigo Durán, CENIA's general manager, acknowledges this is a long-term goal that could take a decade to fully realize. For now, he believes the project's greatest contribution "will be demonstrating that we—Latin America and the Caribbean—have the capabilities, the talent" to undertake such a groundbreaking initiative. It's a strategic first step toward closing the AI gap and building a digital future that truly reflects the region it aims to serve.
