Many of the world’s most popular artificial intelligence (AI) tools are powered by biased programs developed by companies like OpenAI and Meta, according to a new study released by the United Nations Educational, Scientific and Cultural Organization (UNESCO) on March 7.
The biggest companies in the AI race are using massive amounts of data from the internet to train their algorithms, known as large language models (LLMs). UNESCO tested Meta’s Llama 2 algorithm and OpenAI’s GPT-2 and GPT-3.5, used in the popular chat app ChatGPT. The tests found that each showed clear evidence of bias against women.
The results showed that the texts related to female names generated by these tools often contained words such as “house”, “family” or “children”, while the texts related to male names often contained words such as “business”, “salary” or “career”; men were often described with jobs such as teachers, lawyers, doctors; while women were often associated with jobs such as cooking, housework, etc.
DO CAO
Source
Comment (0)