IA: UNESCO warns that ChatGPT propagates sexism – 03/07/2024 – Tech

Meta and OpenAI’s language models, which serve as the basis for their generative artificial intelligence tools, have propagated sexist prejudices, warns a study published by UNESCO this Thursday (7), one day before International Women’s Day.

OpenAI’s GPT 2 and GPT 3.5 models — this one incorporated into the free version of ChatGPT, as well as Meta’s Llama 2 — reveal “unequivocal evidence of prejudice against women”, said the UN agency in a statement.

“Discriminations in the real world are reflected in the digital sphere and are also amplified” in these tools, said Tawfik Jelassi, UNESCO’s deputy director-general for communication and information.

According to the study, carried out from August 2023 to March 2024, female names in these language models are more associated with words such as “house”, “family” and “children” and male names, with “commerce”, “salary ” and “career”.

Researchers asked these interfaces to produce stories about people of different backgrounds and genders. The results showed that stories about “people from minority cultures are often repetitive and based on stereotypes.”

And so, the English man was repeatedly presented as a teacher, driver or bank clerk while the English woman, in almost a third of the texts generated, appeared as a prostitute, model or waitress.

These companies “cannot represent all their users”, said Leona Verdadero, an expert in digital policies and digital transformation at UNESCO.

“Every day more people use language models at work, at school and at home,” said Audrey Azoulay, director general of UNESCO, highlighting that these tools have the power to “subtly shape” people’s perceptions.

“Even subtle gender biases in your content can significantly magnify real-world inequalities,” warned Azoulay.

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

In teams that work with artificial intelligence, globally, only 22% are women, according to figures from the World Economic Forum.

The UN body also calls on governments to regulate the sector more intensively to implement “ethical artificial intelligence”.