Evaluation of AI Ethics Tools in Language Models

• Grant number: 24/23118-1
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: June 01, 2025
• End date: August 31, 2028
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Hélio Pedrini
• Grantee: Jhessica Victoria Santos da Silva
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Associated research grant: 23/12865-8 - Horus: artificial intelligence techniques to detect and forestall synthetic realities, AP.TEM

Abstract

In Artificial Intelligence (AI), language models have gained an important place due to the significant advances made in various fields of knowledge and the recent popularization of systems capable of simulating realistic conversations with human beings through the generation of texts. Because of their impact on society, developing and deploying these language models must be done responsibly, with attention to their negative impacts and possible harms. In the past few years, there has been a rise in the publication of AI Ethics Tools (AIETs), with the aim of helping developers, companies, governments, and other interested parties to establish trust, transparency, and responsibility with their technologies. In our previous work, we proposed a methodology for selecting and evaluating AIETs in language models. We carried out a literature survey and defined criteria to filter all the AIETs found. Through interviews with developers of language models in Portuguese, we quantitatively evaluated 4 AIETs. The evaluation considered the developers' perceptions of the AIETs' use and quality in helping to identify ethical considerations about language models and whether the concerns raised correspond to the same ones present in the literature on ethical impacts of language models. This research project proposes to extend our previous work by carrying out a systematic qualitative analysis of the previously conducted interviews. The aim of this analysis is to gain deeper insights into the impacts of language models developed for the Portuguese language and how developers understand these impacts, while also considering the nuances of the conversations during the interviews and the logic for mapping the model's risks. Following this analysis, the aim is to identify the gaps in the AIETs evaluated and thus propose a new AIET exclusively for language models.