A Survey of Methods for Assessing Trustworthiness in Artificial Intelligence Systems
Plachimowicz, Dagmara (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025060420358
https://urn.fi/URN:NBN:fi:amk-2025060420358
Tiivistelmä
This thesis investigates practical methods for evaluating and validating the technical and safety aspects of trustworthy artificial intelligence (AI) systems, with a focus on their applicability within the MANOLO project—a European initiative aiming to deploy efficient, trustworthy AI in resource-constrained environments such as industrial robots, healthcare devices, and telecom infrastructure. The development task was to identify and analyse existing trustworthy AI assessment approaches and propose a technically focused framework that complements MANOLO’s adoption of the Z-Inspection® process, which currently lacks detailed tools for evaluating technical robustness.
The methodology followed is a systematic literature review embodied within the PRISMA reporting guidelines, which involved a comprehensive screening and quality assessment of peer-reviewed primary studies published between 2018 and early 2025.
The review identified five categories of trustworthy AI assessment approaches, including checklist-driven frameworks, formal verification tools, audit-based models, statistical methods, and hybrid evaluation schemes. Findings reveal that while various methods exist, many are either too general, lack operational clarity, or do not cover technical trustworthiness in sufficient depth. Several studies demonstrated potential to bridge this gap, especially those offering domain-specific implementation procedures and metrics linked to robustness, reliability, or safety.
The results were analysed with a focus on practical applicability to MANOLO’s technical use cases. The study concludes by proposing a hybrid framework that integrates technically grounded assessment procedures into the Z-Inspection® process. This framework facilitates the evaluation of AI systems’ operational performance and resilience, making it particularly suited for deployment in edge-based and safety-critical contexts.
The methodology followed is a systematic literature review embodied within the PRISMA reporting guidelines, which involved a comprehensive screening and quality assessment of peer-reviewed primary studies published between 2018 and early 2025.
The review identified five categories of trustworthy AI assessment approaches, including checklist-driven frameworks, formal verification tools, audit-based models, statistical methods, and hybrid evaluation schemes. Findings reveal that while various methods exist, many are either too general, lack operational clarity, or do not cover technical trustworthiness in sufficient depth. Several studies demonstrated potential to bridge this gap, especially those offering domain-specific implementation procedures and metrics linked to robustness, reliability, or safety.
The results were analysed with a focus on practical applicability to MANOLO’s technical use cases. The study concludes by proposing a hybrid framework that integrates technically grounded assessment procedures into the Z-Inspection® process. This framework facilitates the evaluation of AI systems’ operational performance and resilience, making it particularly suited for deployment in edge-based and safety-critical contexts.