Proceedings of The 6th Global Conference on Women’s Studies
Year: 2024
DOI:
[PDF]
Detecting and reducing gender bias in AI chatbots: the Lovelace project
Irene Carrillo, César Fernández, M. Asunción Vicente, Mercedes Guilabert, Alicia Sánchez, Eva Gil, Almudena Arroyo, María Calderón, M. Concepción Carratalá, Adriana López, Ángela Coves, Elisa Chilet, Sergio Valero, Carolina Senabre
ABSTRACT:
As part of the Lovelace project (Development of methods for evaluating gender bias in AI chatbots, https://lovelace.umh.es), we have developed a protocol that allows for the evaluation of gender bias in various generative artificial intelligence platforms and enables the establishment of measures for its correction through the addition of ad-hoc designed contexts. Gender bias detection is carried out using Microsoft’s open-source GenBit software (https://github.com/microsoft/responsible-ai-toolbox-genbit), and specifically, we use the genbit_score metric (average of the absolute bias conditional ratios per token), as it is one of the most widely used standards in the literature. So far, we have worked with the generative AIs ChatGPT and Mistral through their respective APIs. The test battery consists of a set of requests for generative AIs where the results are likely to show significant gender bias or magnify female and male roles. To ensure the reliability of the results obtained, all requests are repeated 10 times with each generative AI and with each context, so that the natural variability of responses offered by generative AIs does not alter the results. The software platform developed for the project is based on cloud spreadsheets and scripts in PHP and Python. Both the software, the data, and the results obtained are available to the scientific community for further study and analysis. The results obtained so far allow us to verify that the addition of specifically designed contexts helps to limit gender biases in the responses of both generative AIs.
keywords: AI context; ChatGPT; GenBit; generative AI; Mistral