© Pexels/Matheus Bertelli
On this International Women's Rights Day, the study Prejudice against women and girls in UNESCO's major language models is timely. While artificial intelligence is booming and increasingly interfering in our daily lives, the United Nations Educational, Scientific and Cultural Organization highlights a less than stellar side of ChatGPT et al…
Indeed, the UNESCO study published on Thursday March 7, 2024 demonstrates that generative artificial intelligence has integrated a large quantity of prejudices against women. Unfortunately, this isn't really surprising.
Women in the kitchen for ChatGPT ?
UNESCO conducted the survey between August 2023 and early March 2024. The global body questioned the natural language processing tools behind the most widespread generative AI platforms in the world. actual hour. GPT-3.5 and GPT-2 from OpenAI and Llama 2 from Meta were thus analyzed with interest. It's indisputable: AI exhibits significant gender bias.
So, when asked to generate content, these platforms frequently associate women with the words “home”, “family”, “love” or even “garden”. On the other hand, AI more readily uses words like “career”, “salary” and “company” when talking about men. Likewise, artificial intelligence more easily sends women to less valued roles such as “cook”, “domestic” or even “prostitute”. Men have the right to more diverse and high-status jobs. In stories generated by AI, men benefit from more prestigious roles, such as “engineer”, “doctor” or “teacher”.
Is this really surprising? No. Indeed, to be able to generate coherent and accurate content, tools like ChatGPT had to ingest a lot of textual data. These come from multiple sources on the Internet, such as books, articles, websites, online discussions. The problem is that sexist biases and gender prejudices are legion on our dear Internet. QED.
In addition to the latent sexism of AI, there are also many racist stereotypes and homophobic attitudes. As Tawfik Jelassi, UNESCO Assistant Director-General for Communication and Information, explains, “the discriminations of the world are not only reflected in the digital sphere, they are also amplified there”.
To combat these cognitive biases, UNESCO has shared several recommendations. Such as diversifying recruitment in machine learning companies, financially encouraging female entrepreneurship or investing in targeted programs to increase the opportunities for girls and women to participate in scientific disciplines, technological, or engineering.
Good news: the open and transparent nature of Open Source LLMs like Llama 2 and GPT-2 can make it possible to take the opposite approach to address and mitigate these prejudices. To achieve this, UNESCO advocates increased collaboration within the global scientific community. But this is more complicated with more closed models, like GPT-3.5, GPT-4 and Gemini (Google).
Nevertheless, when asked, ChatGPT claims not to be sexist. Well, everything is fine.
Screenshot of ChatGPT-3.5 © Presse- lemon
📍 To not miss any news from Presse-citron, follow us on Google News and WhatsApp.
[ ]