Vivek Badrinath Column | AI Language Bias And Economic Risks Explained

Vivek Badrinath Director General of Global System for Mobile Communications

Much has been said about the security and military risks of AI. But the cultural and economic threats posed by Large Language Models (LLMs), trained in a small number of languages ​​and owned by a few multinational companies, have been largely ignored. The market dominance of these LLMs puts developing countries with marginalized languages ​​at a significant disadvantage.

Since OpenAI released ChatGPT in November 2022, industry-leaders and policy-makers have praised AI’s capabilities to increase productivity. But its benefits do not reach everyone equally. Many developing economies do not have access to these AI systems due to their language biases.

Most LLMs are trained in only about 100 of the more than 7,000 languages ​​spoken in the world, and almost all the data used comes from English, including Chinese, Spanish, French and German.

But only 400 million people in the world are native English speakers and only 25% of the global population speaks one of the above ‘Big Five’ as their first language. This imbalance may lead to mere cultural and economic isolation in the short run, but in the long run it can hinder the development of many countries and reduce linguistic diversity.

There is also an irony that the speed with which rich countries have adopted AI has made it difficult for users in developing countries to access the Internet and benefit from digital tools. The demand for AI-powered chips has also increased the cost of smartphones, putting them out of reach for millions of people.

LLM should better reflect local languages ​​and cultures. Otherwise communities risk losing their unique vocabularies. These terminologies are shaped by historical events, mass migration, and political legacies. The lack of linguistic diversity in AI systems will also slow down their adoption.

Digital financial services in Peru must communicate in colloquial language to be relevant to consumers. To make online healthcare work well in Vietnam, it will need the clarity that only a local language model can provide.

Mobile network operators (MNOs) are in a unique position to create local language models due to their large number of developers and data-processing capabilities. They are also licensed and trusted government partners.

This is important because a major hurdle in creating LLMs for marginalized languages ​​is the lack of training-data. The solution lies in the vast amounts of data held by governments, which is usually sensitive and confidential.

For example, Kyivstar, Ukraine’s largest MNO, formed a partnership with the Ministry of Digital Transformation to develop the first national LLM there, which was adapted for Ukrainian as well as other languages ​​spoken there—Russian, Bulgarian, and Crimean Tatar.

The Global System for Mobile Communications Association—of which I am Director General—is pursuing a similar strategy in much of sub-Saharan Africa, including Nigeria, Madagascar, Togo, and the Republic of Congo, through the African AI Language Models project.

The initiative is bringing together the continent’s six MNOs – Airtel, Axion Telecom, Ethio Telecom, MTN, Orange, Vodacom and other telecom companies, public and private sector individuals, academics and civil society groups to create an LLM for African languages. We need to take a broader view of the uses of AI.

The world is now moving towards agentic AI. Digital services like banking, healthcare, education are increasingly adopting AI to assist users. In such a situation, it would be very wrong if any country lags behind on the basis of language.

(@ProjectSyndicate)

Make Dainik Bhaskar your favorite source on Google ➔