Chile will on Tuesday launch Latam-GPT, an open-source artificial intelligence model designed to counter cultural bias embedded in a largely US-dominated global AI industry.
Developed by the Chilean National Centre for Artificial Intelligence (CENIA), Latam-GPT has been trained on millions of data points sourced from across Latin America, with the aim of better reflecting the region’s cultural, linguistic and social diversity.
The project seeks to address what Chile’s Science Minister Aldo Valle described as entrenched “prejudices” in existing AI systems, which often portray Latin America as culturally uniform.
“Models developed elsewhere do include data from Latin America, but it represents a very small proportion,” said Álvaro Soto, director of CENIA.
That imbalance is reflected in how global AI systems depict the region. In ChatGPT, for example, a typical Chilean man is often portrayed wearing a poncho against an Andean backdrop.
“Latin America cannot simply be a passive user or recipient of artificial intelligence systems,” Valle argued. “That could lead to the loss of a significant part of our traditions.”
Unlike closed generative models such as ChatGPT or Google’s Gemini, Latam-GPT is fully open-source, allowing developers to adapt and customise the software for local needs.
The project draws on contributions from universities, foundations, libraries, government bodies and civil society organisations across the continent, including Argentina, Brazil, Chile, Colombia, Ecuador, Mexico, Peru and Uruguay.
Indigenous content and regional languages
US technology firms continue to dominate the global AI race, with low-cost Chinese models rapidly gaining ground and Europe trailing behind. However, other regions are increasingly recognising the importance of developing public AI systems aligned with local cultural norms and safety standards.
In 2023, researchers in Singapore released the open-source SEA-LION model for Southeast Asian languages, while in Kenya the UlizaLLama large language model has been used to provide healthcare support for Swahili-speaking expectant mothers.
Latam-GPT has been trained on more than eight terabytes of data, equivalent to millions of books. Remarkably, it was developed on a budget of just $550,000, funded primarily by the Development Bank of Latin America (CAF) and CENIA’s own resources.
An initial version was trained using Amazon Web Services, but future development will take place on a supercomputer at the University of Tarapacá in northern Chile.
At present, the model is trained mainly in Spanish and Portuguese, though developers plan to incorporate content in Indigenous Latin American languages.
Slang, idioms and local nuance
Latam-GPT will be made available free of charge to companies and public institutions seeking to develop applications tailored specifically to Latin America, Soto said.
He cited potential uses in healthcare, including hospital systems grappling with logistical challenges and inefficient use of medical resources.
Given its modest funding, Latam-GPT has “no chance” of competing directly with major global AI models, according to Alejandro Barros, a professor of industrial engineering at the University of Chile.
Nevertheless, it has already attracted interest from local entrepreneurs. Roberto Musso, founder of Chilean digital firm Digevo, said his company plans to use Latam-GPT to build customer service platforms for airlines and retailers.
Clients, Musso said, are keen for users to be able to communicate and receive responses in genuinely local language.
Latam-GPT, he added, can recognise “slang, idioms and even speech patterns”, while avoiding regional biases commonly found in other AI systems.
