Towards Resilient and Responsible AI: A Scoping Review of Barriers and Enablers

Abstract

Artificial intelligence (AI) is developing constantly, and its overall impact remains uncertain. Therefore, it is important to examine how AI systems can be developed and adopted in ways that enhance their resilience while mitigating potential risks and unintended consequences. Resilience is a multidisciplinary concept, which is often related to the capability to anticipate and withstand disruptions as well as adapt to changes, while also preserving core structures and functions of the system. In recent years, the importance of resilience has been emphasized by various institutional players. For example, the European Union (EU) has created the concept of Industry 5.0, which promotes human-centric, sustainable, and resilient systems. However, it is still uncertain what resilience refers to in the context of AI and how it can be effectively incorporated into AI systems. At the same time, AI systems are expected to fulfil ethical and regulatory expectations. From this perspective, Responsible AI (RAI) offers a potential framework for integrating ethical requirements with technical and organizational dimensions of AI system resilience. RAI is as concept that promotes the ethical and responsible development of AI systems through various principles, such as transparency and fairness. These principles can support resilience by improving robustness, trustworthiness, and adaptability. Furthermore, a range of guidelines, regulatory frameworks, and governance mechanisms have been developed to facilitate the practical implementation of RAI. Nevertheless, the relationship between resilience and RAI remains underexplored.

This research investigates how the academic literature conceptualizes AI system resilience, referring to the key characteristics that define resilient AI systems, and how RAI is linked to it. This is done by conducting a scoping review that is focused on a variable-focused approach by identifying the barriers that may cause disruptions to the AI system and enablers that can help the AI system to anticipate, cope with, and adapt to the changes caused by the disruptions, while also preserving the core functions and qualities of the system. The findings of this research are based on 52 articles and conference papers published between 2020 and 2025. The research papers were chosen based on predefined inclusion and exclusion criteria as well as systematic and transparent review and reporting methods. The literature was retrieved from three databases: Scopus, Web of Science Core Collection, and IEEE Xplore. The reporting of this research followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) and the flow diagram of the PRISMA 2020 Statement. The barriers and enablers were identified through thematic analysis that included coding of the important themes and concept mapping. Moreover, the ethical AI principles were utilized to investigate the RAI perspective.

First, AI system resilience is conceptualized according to three themes: performance continuity, adaptation and recovery, and system functionality. Subsequently, the barriers are presented according to six groups: hardware and model issues, data issues, network and communication issues, human-related and organizational challenges, and ethical and regulatory issues. Similarly, the enablers are presented according to six groups: data quality and management, clear requirements and objectives, resilient model architecture, versatile training and learning methods, efficient fault detection and diagnosis, and ethical and organizational enablers. Overall, the findings suggest that AI system resilience should be understood as a holistic system property, which considers technical characteristics, human users, and other stakeholders involved in the system operations. AI system resilience encompasses the ability of the system to anticipate, cope with, and adapt to disturbances, while preserving the core functions of the system and learning from setbacks. The identified barriers and enablers were related to the principles of transparency, justice and fairness, non-maleficence, responsibility, privacy, trust, and sustainability. In relation to AI system resilience, RAI functions as a hindering and an enabling factor, and users and organizations have an essential role in defining how it is manifested in practice. To conclude, this research unifies the literature related to AI system resilience and presents various themes that challenge and support the development and utilization of resilient AI systems from the perspective of RAI.