⚡ Quick Summary
This paper is a systematic literature review by Louise McCormack and Malika Bendechache of the ADAPT Research Centre, School of Computer Science, University of Galway. It carries the arXiv identifier arXiv:2410.17281v1 and states that the work has been accepted for publication in AI and Ethics. It reviews existing evaluation criteria and metrics for Trustworthy Artificial Intelligence (TAI) through the seven EU principles, from the AI HLEG ethics guidelines and ALTAI to the EU AI Act and ISO/IEC 42001.
The review follows an established systematic review process, reports 971 papers returned by a Google Scholar search string and draws on 63 publications contributing to the core findings. It answers five research questions on standards initiatives, current metrics, differences between principles, multi-principle scoring and the obstacles to a scoring and assessment system. Its main deliverable is a proposed classification of evaluation criteria for each principle, including fairness classes that extend Caton and Haas with Intersectional Fairness, Complex Fairness and Inclusive Design and Participation, and classifications for Data, Model and Outcome Transparency.
The paper concludes that criteria and benchmarks remain unstandardised and that the principles are interrelated, so trade-offs are unavoidable. It argues for use case-specific metrics agreed at sector level and illustrates the approach with a mortgage lending case study.
🧩 What’s Covered
- Background and EU framing: Trustworthy AI is defined as the ethical considerations for building and using AI, traced from AI4People's five metrics (2018) and the AI HLEG ethics guidelines (2019) to ALTAI (2020); the seven EU principles are set out individually and supported by a table of working definitions, alongside the EU AI Act's documentation duties for high-risk systems and ISO/IEC 42001.
- Methodology and review results: a three-phase systematic review with five research questions, the Google Scholar search string, 971 papers returned, the inclusion and exclusion criteria in Table 1, 63 publications contributing to the core findings, the data extraction fields, and the rise in publications from 2020.
- Fairness: the most mature principle, with group fairness metrics (parity, confusion matrix, calibration and score-based), individual and counterfactual metrics, and three newly proposed classes: Intersectional Fairness, Complex Fairness (Alternate World Index, Fairbench, CLEAM, FairEvalLLM) and Inclusive Design and Participation.
- Transparency: Data, Model and Outcome Transparency, drawing on the Foundation Model Transparency Index, Fehr et al.'s five assessment areas, SHAP, LIME and PLENARY, ISO/IEC 42001's data provisions and ISO/IEC 25059.
- Human agency and oversight: human control (user control, feedback quality, training, the ability to stop the system) and the human-AI relationship (human-in-the-loop involvement, expert oversight, user satisfaction and perceived trust).
- Privacy and data governance: differential privacy and the epsilon parameter, data leakage measurement such as the ML Privacy Meter, and criteria for data collection, processing, compliance and consistency, referencing ISO/IEC 38507.
- Technical robustness and safety: adversarial and side-channel attack testing, the Risk Priority Number and Overall Criticality Number from TAI-PRM, robustness metrics (MCC, VIF, TSVR, CSVP, Shapiro-Wilk, Breusch-Pagan) and ISO/IEC TR 24029-1 and -2.
- Accountability, societal well-being and the case study: auditability and risk management (traceability, AI Cards, ISO 37002), societal impact and sustainability including FLOPs, trade-offs between principles, and the mortgage lending example in Table 3.
💡 Why it matters?
The review addresses a practical gap: organisations that must select, justify and document evaluation metrics for AI systems, as the EU AI Act requires high-risk systems to explain their performance metrics and their suitability. It maps what can be measured today for each principle and where only qualitative, manually assessed criteria exist, helping governance teams separate automated evidence from documentation-based assurance. By naming positive correlations and trade-offs between principles, it supports the documented trade-off decisions the Act expects, and it anchors its classification in ALTAI, the AI Act and ISO/IEC 42001 rather than proposing a separate regime.
❓ What’s Missing
Three principles — human agency and oversight, accountability, and societal and environmental well-being — have no quantifiable metrics in the reviewed literature, and outcome transparency has no dedicated metric; for these areas the classification lists evaluation considerations rather than measurable criteria. The authors state that research into metric selection is thin and that no consensus exists on thresholds. Evaluation methods (automated, semi-automated, manual, conceptual) are explicitly out of scope, as are non-English publications. The classification is high-level and informed mainly by ALTAI, the single worked case study is mortgage lending, and the proposed classes are not empirically validated here.
👥 Best For
Compliance and AI governance leads who must select and justify metrics for high-risk systems; risk, audit and assurance teams building trustworthy AI assessments around ALTAI and ISO/IEC 42001; machine learning and data science teams choosing fairness, robustness or privacy metrics; and researchers extending TAI evaluation criteria and benchmarks.
📄 Source Details
A Comprehensive Survey and Classification of Evaluation Criteria for Trustworthy Artificial Intelligence by Louise McCormack and Malika Bendechache, ADAPT Research Centre, School of Computer Science, University of Galway. The document is an arXiv preprint carrying the identifier arXiv:2410.17281v1 dated 10 October 2024, and states that it has been accepted for publication in AI and Ethics. The supplied extraction covers all 39 pages; the article body is numbered to page 30, followed by references. No URL for the document itself is printed.
