Parlez-vous AI? Francophone scholars call for multilingual models

Francophone scholars are sounding the alarm about the dominance of English in artificial intelligence (AI) development, warning about the potential risks to “global cultural diversity” and “national sovereignty”. 

In an open letter published by the Future of Life Institute on Friday, the experts said there is a lack of linguistic and cultural diversity in AI models and safety assessments.

They argue these shortcomings “pose a threat to the national sovereignty of the states in which they are distributed and to the safety of users”.

The experts are calling on global leaders and tech giants to prioritise multilingualism in AI development, viewing it not as an afterthought, but as a fundamental aspect of creating safe, equitable, and truly global AI. 

In January, researchers at Brown University discovered that safeguards designed to prevent OpenAI’s GPT-4 from answering dangerous prompts failed when it received requests in languages such as Zulu or Scots Gaelic. This allowed researchers to get AI-generated answers on how to build a homemade bomb. 

This vulnerability was made possible as the instruction was made to the AI in languages that are mostly absent from its training data. The translations were done with Google Translate. 

The letter’s publication comes the same day as the Francophonie summit begins and as France prepares to host the AI Action Summit in February. 

France has made no secret that it wants to become an AI hub, with the government investing in the technology and generative AI start-up superstars MistralAI and H Company.

However, the major tech companies release and develop their AI models in English. 

Despite AI models being available in many other languages, many local languages are not included and datasets are usually trained in English.

The scholars said AI’s English language-centric approach may lead to critical misinterpretations and flawed decision-making in diverse linguistic contexts. 

Signatories of the letter include Mohamed Farahat, vice-chairman of the multistakeholder advisory group (MAG) of the African Internet Governance Forum (AFIGF) and representatives which include the Swiss Federal Institute of Technology Lausanne, and the Tunisian Ministry of Education. 

“The letter calls for international standards to ensure that a system is safe and beneficial for all users, and it addresses the critical need for multilingual and multicultural evaluations,” Imane Bello, AI safety summit lead at the Future of Life Institute, told Euronews Next. 

However, she said this was also important for English-speaking countries.

“If considerations related to multi-linguistic AI safety are not embedded then the current and ongoing work around the science of evaluations might miss a very important part,” she said. 

ADVERTISEMENT

“The second thing is not having this particular important section can lead to harm in vulnerable communities,” she added.