Latam-GPT: A Latin American AI to combat US-centric bias

Move over ChatGPT -- on Tuesday (10 February), Chile will launch Latam-GPT, an open-source artificial intelligence model designed to combat biases built by the primarily US-centric industry.

Developed by the Chilean National Center for Artificial Intelligence (CENIA), Latam-GPT uses millions of data points collected in Latin America to showcase the continent's cultural diversity.

The program seeks to combat "prejudices" and prevent depictions of Latin America always "looking the same," Chilean Science Minister Aldo Valle told AFP.

"The models developed in other parts of the world do have data from Latin America, but it represent a fairly small proportion," CENIA director Alvaro Soto noted.

The low level of local input is reflected in the depictions of Latin Americans by major AI models.

For ChatGPT, a typical Chilean man is a person wearing a poncho with the Andes in the background.

Latin America "cannot simply be a passive user or recipient of artificial intelligence systems. That could result in the loss of a significant part of our traditions," Valle argued.

Unlike closed generative models like ChatGPT or Google's Gemini, Latam-GPT is an open model that can be used by programmers to customise parts of the software to suit their needs.

Contributions to the project were provided by universities, foundations, libraries, government entities, and civil society organisations from across the continent, including Argentina, Brazil, Chile, Colombia, Ecuador, Mexico, Peru and Uruguay.

Indigenous content

Major US tech companies dominate the global AI race, with low-cost Chinese models rapidly gaining ground and Europe lagging in third.

But other regions are also coming onto the importance of developing public AI models that respect their cultural norms and safety standards.

In 2023, Singapore researchers released the open-source Southeast Asian Languages in One Network, or SEA-LION model, whereas in Kenya, the UlizaLLama LLM provides health services for Swahili-speaking expectant mothers.

Latam-GPT has been trained on more than eight terabytes of data, equivalent to millions of books.

It was developed for a mere $550,000, primarily from the Development Bank of Latin America (CAF) and Cenia's own resources.

A first version was developed for the Amazon Web Services cloud, but in the future Latam-GPT will be trained on a supercomputer at the University of Tarapaca in northern Chile.

At the moment, it is trained mainly in Spanish and Portuguese content, although its developers plan to also incorporate material in Indigenous Latin American languages.

Slang and sayings

Latam-GPT will be available free of charge to companies and public institutions to develop applications that are "more specific" to Latin America," said Soto, the CENIA director.

He gave as an example the developments of applications for hospitals "with logistical problems or issues with the use of medical resources."

Its tiny budget means it has "no chance" of competing against the major AI models, Alejandro Barros, a professor in the Department of Industrial Engineering at the University of Chile told AFP.

But it has already won over Chilean serial digital entrepreneur Roberto Musso.

Musso's company Digevo plans to use Latam-GPT to develop customer service programs for airlines or retailers.

Musso said his clients were "very interested in having their users express themselves and receive responses in the local language."

Latam-GPT, he said, provides the ability to recognise "slang, idioms, and even speech rate" and avoid certain biases with regard to the region that can arise in other AI models.