PERIAC: um processo para elicitação de requisitos para inteligência artificial confiável

Abstract

Artificial Intelligence (AI) has brought new challenges and requirements to Software Engineering (SE) projects, driving the need to ensure reliable AI systems that meet both technical and ethical requirements. Government bodies and the scientific community have been working to promote the development of trustworthy AI, emphasizing the importance of techniques that enable the active participation of all stakeholders in the Requirements Engineering (RE) process, including end users. This dissertation proposes the development of a process for eliciting Non-Functional Requirements (NFRs) aimed at reliable AI, involving different roles in the development of these systems. The proposed process, named the Process for Elicitation of Requirements for Reliable AI (PERIAC), encompasses the fundamental requirements identified in the literature, such as Fairness, Explainability, Accountability, Privacy, and Acceptance. These requirements ensure that AI systems are designed robustly, guaranteeing both technical performance and ethical compliance. The Brainwriting technique was employed to facilitate the collaborative generation of ideas among the various stakeholders involved in AI development. This technique proved effective in fostering multidisciplinary participation and including diverse perspectives in the requirements definition process. During the application of PERIAC, a set of questions directly drawn from the literature was developed, as well as a set of cards representing the five essential principles of trustworthy AI: Fairness, Explainability, Accountability, Privacy, and Acceptance. The process was applied in a case study, allowing the elicitation of both functional and non-functional requirements. The application of PERIAC demonstrated that the process, in addition to fulfilling its goal of eliciting NFRs, was also effective in generating Functional Requirements (FRs), broadening the initial scope of the research. This finding revealed that the reliability of AI systems depends on the intersection of FRs and NFRs, making the system reliable as a whole.