About 50 research groups that plan to build AI applications on ChatGPT and other large language models to help social and healthcare causes in their communities were granted $100,000 each from the Gates Foundation
Getty
Researchers selected for Gates Foundations’ new program plan to build AI-based applications to address crucial healthcare and social causes in their communities.
The Bill & Melinda Gates Foundation on Wednesday announced 48 recipients of a $5 million program to develop AI-based applications built on large language models that target pressing problems in low-and-middle-income countries.
The grantees, who will each be awarded $100,000, are working on issues that span a broad spectrum, from researchers developing a ChatGPT-based chatbot to create and manage detailed electronic medical records for maternal healthcare workers in Pakistan to an entrepreneur working on an AI tutoring tool to offer specialized education to students in Kenya.
While a majority of grantees are aiming to test the use of generative AI for healthcare issues like HIV risk assessment, prenatal care and antibiotic prescriptions, several are focused on applying the technology to other local issues. For instance, one group of scientists in Uganda plan to use the funds to build a ChatGPT-based application to provide farmers with information about crop diseases; as part of the project, the scientists are planning to build a dataset in the native Luganda language. In Vietnam, a researcher is creating a chatbot to provide advice to residents in an area affected by saltwater intrusion by fine tuning GPT-4 with data in Vietnamese. In Brazil, a non profit organization plans to use LLMs to develop a support bot for psychologists and lawyers helping women who have faced gender-based violence.
Right now, the vast majority of the big AI companies are located in the Global North. This initiative’s goal is to encourage the development of generative AI across the world, so that more people can benefit from the technology.
“Too often, advances in technology deliver uneven benefits in many parts of the world due to existing patterns of discrimination, inequality, and bias,” said Kenyan computer scientist Juliana Rotich, who is part of the foundation’s AI safety committee. “Most of the tools being developed in the Global North are using data from lower-resourced regions that is often incomplete or inaccurate.”
Chatbot applications that require users to enter text-based prompts can exclude a large part of the population, such as non-English speakers and those without smartphones. That’s why some researchers also plan to develop a feature that converts a person’s voice (in a local language) to text to make generative AI more accessible.
They also have to contend with the faults of ChatGPT and the like, which are trained on billions of parameters of unfiltered public data and struggle with factual inaccuracies and racial and gender biases. To address this concern, the foundation has created a support hub of global AI experts to guide the grantees in assessing potential risks.
Over two weeks, Brey’s team of 80 reviewers received about 1300 proposals from researchers, nonprofit organizations and private companies across 103 countries. These applications were judged on the basis of several criteria, including that the work must be done in a low-and-middle-income country, must be focused on a critical societal issue and that they must use a large language model for their application. The last criteria was key because the program’s goal is to gauge what practical problems LLMs would present for users in developing nations, such as how accessible these tools are, said Zameer Brey, an interim deputy director leading the foundation’s AI efforts.
The selected recipients will have three months to complete their projects, for which they are mainly using and fine tuning OpenAI’s GPT-4 and GPT 3.5, with a handful of projects using LLMs like Google’s LaMDA and Bert and a text-to-text model trained on 100 languages called mT5.
Through the new program, the Gates Foundation hopes to build an “evidence base” of generative AI use cases, roadblocks and learnings, while determining how AI can find its place in low-income communities, Brey said.
“I think, as a foundation, we acknowledge the hype, but we want to channel the hype towards providing good evidence for decision making and implementation,” he said.