Trustworthy AI in Practice: A Comprehensive Review of Human Oversight and Human-in-the-Loop Approaches
Year: 2025
Type: preprint
Study design: Systematic Review
Abstract: The widespread adoption of Artificial Intelligence (AI) across various domains has underscored the critical need for a framework that ensures its safe and responsible use. This paper presents a comprehensive review of human oversight and human-in-the-loop approaches, analyzing their role as a fundamental requirement for achieving trustworthy AI. While AI regulatory frameworks, such as the EU AI Act, mandate human oversight to mitigate risks and preserve human autonomy, the practical implementation remains ambiguous. To address this, our analysis draws from a systematic review of 203 publications spanning 2014-2024, examining the intersection of trustworthy AI and human oversight across major academic databases. We employ a lifecycleoriented perspective that maps regulatory requirements from the EU AI Act to technical implementations, providing empirical evidence from diverse domains including healthcare, finance, autonomous vehicles, and judicial systems. The survey distinguishes itself by positioning human oversight as both a compliance obligation and an enabling mechanism for other trustworthiness dimensions, offering a novel taxonomy that connects oversight approaches to specific AI development phases and trust requirements. Our investigation reveals that human oversight operates through three complementary mechanisms with distinct intervention points and control functions. HITL approaches enable continuous human participation in decision-making processes, particularly crucial in high-stakes applications such as medical diagnostics and judicial decisions. HOTL mechanisms facilitate supervisory monitoring with selective intervention capabilities, exemplified in autonomous vehicle operation and fraud detection systems. HIC approaches maintain strategic authority over AI governance, ensuring alignment with ethical standards and regulatory compliance across the system lifecycle. In addition, these mechanisms translate abstract regulatory mandates into concrete technical implementations, utilizing components such as interactive interfaces, decision support systems, real-time monitoring dashboards, and documentation frameworks. This work makes four principal contributions to the field of trustworthy AI. First, we establish human oversight as a meta-requirement that operationalizes and validates other trustworthiness attributes, fairness through bias detection and mitigation, robustness through adversarial defense validation, explainability through interpretation facilitation, and privacy through data governance enforcement. Second, we provide a structured taxonomy mapping oversight mechanisms to AI lifecycle stages, demonstrating how different oversight approaches address stage-specific risks and requirements. Third, we identify and analyze critical implementation challenges, including operational ambiguity in regulatory frameworks, highlight complex trade-offs between oversight mechanisms, alongside other related issues. Fourth, we propose actionable future research directions addressing harmonization of global governance frameworks, standardization of trustworthiness measurement, development of adaptive hybrid oversight models, and integration of emerging technologies.
Institution: University of Tartu
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: trustworthy AI, human oversight, human-in-the-loop, EU AI Act, AI regulation, AI lifecycle, AI governance, regulatory compliance, decision support system, real-time monitoring, explainability, privacy, autonomous vehicles, judicial decision-making, artificial intelligence in healthcare, fraud detection
Open Access status: green
Funders: European Commission, Eesti Teadusagentuur
OpenAlex ID: W4415450413
