Large language models developed by Meta and OpenAI, which serve as the basis for their generative artificial intelligence tools, transmit sexist biases, warns a study published by UNESCO on Thursday, ahead of International Women’s Day, AFP and Agerpres reported.

Models GPT 2 and GPT 3.5 from OpenAI, the latter underpinning the free version of ChatGPT, as well as Llama 2 from competitor Meta, show a “clear bias against women,” the UN agency said in a statement.

“Discrimination in the real world is not only reflected in the digital sphere, but also exacerbated there,” emphasized Tawfik Jelassi, UNESCO’s Deputy Director-General for Communications and Information.

According to the study, conducted between August 2023 and March 2024, these linguistic patterns are more likely to associate feminine nouns with words like “house,” “family,” and “children,” while masculine nouns are more associated with the words “business”, “salary” and “career”.

The researchers also asked these interfaces to create texts about people of different backgrounds and “genders”. The results showed that texts about “people from minority cultures or about women were often repetitive and based on stereotypes.”

Thus, there is a tendency for an Englishman to be presented as a teacher, driver, or bank clerk, while an Englishwoman will be presented as a prostitute, model, or waitress in at least 30% of the generated texts.

UNESCO warns that AI models could shape people’s perceptions around the world

These companies “are not in a position to represent all their users,” Leona Verdadero, UNESCO’s digital policy and digital transformation specialist, told AFP. As these AI applications are increasingly used by the general public and businesses, they “have the power to shape the perceptions of millions of people,” UNESCO Director-General Audrey Azoulet said.

“Therefore, the presence of even the slightest sexist bias in their content can significantly exacerbate inequality in the real world,” she continued, quoted in the same statement.

To combat these biases, UNESCO recommends that companies in the sector have more diverse engineering teams, especially with more women.

According to data from the World Economic Forum, women make up only 22% of team members working in the field of artificial intelligence around the world, UNESCO reminds.

The same UN agency is also calling on governments to introduce more regulations for the implementation of “ethical artificial intelligence”.