Auditoría de sistemas de IA mediante explicabilidad (AUDITIA-X)

Auditing of AI systems through explainability (AUDITIA-X)

The impact of Artificial Intelligence (AI) in value chains and social missions or challenges is unquestionable. However, the adoption of this technology is not exempt from risks derived from possible biases and inappropriate uses. These risks have been highlighted by the Digital Spain 2025 Plan and recently regulated by the European Union. This project proposes the development of mechanisms that allow the analysis and evaluation of predictive AI systems for the identification of the different biases and risks, and their adaptation to current and future regulatory and ethical frameworks.
The project is rooted in the existing results in the area of eXplicable AI (XAI) for the development of methods that provide transparency to AI systems, specifically those based on machine learning (ML) models. However, it is only very recently - and as a result of emerging regulations- that the concept of explainability has started to be linked to auditing, even though they both have obvious synergies. The appropriate combination of XAI methods, with data analysis and visualization techniques, and complex machine learning model generation processes, can lead to prototypical procedures that allow to comprehensively analyze and evaluate the behavior of ML models.

The main challenge is to address the heterogeneous nature of AI systems and application domains, which makes it difficult to mechanize the auditing and governance procedure of a specific model for a specific domain and dataset. Therefore, the development of generic or theoretical methodologies for the audit of AI systems entails the inherent difficulty of its concrete instantiation to the specific features of such systems, and its adequacy to regulatory and even ethical guidelines where the variability is too high. This project proposes an alternative solution based on the collection of practical and validated use cases that can be reused in auditing similar AI systems. For this purpose, the Case-based Reasoning (CBR) paradigm would be applied, based on the reuse of similar experiences; a field where the applicant research group is a reference and has an experience of more than 20 years.

The development of this project involves the generation of knowledge in different fields that will be of direct benefit to meet the challenge of the implementation of AI in society. On the one hand, concrete results of evaluation and analysis of the risks of AI systems applied to specific problems in high-impact domains such as medicine, cybersecurity, and environmental intelligence, among others, will be developed. On the other hand, a reusable catalog of audit experiences will be made available to the scientific, industrial, and regulatory community to serve as a reference for evaluating and analyzing further AI systems.