The Human Oversight approaches at the forefront of responsible and trustworthy AI, from data-centric adaptive AI-Ops pipelines to Multi-Agent Systems

The paper focuses on human oversight mechanisms in Artificial Intelligence (AI) systems and their pipelines, with particular attention to Human-in-the-Loop (HITL) and Human-on-the-Loop (HOTL) approaches. It explores their main advantages and challenges over fully autonomous systems, from a technical, legal and ethical perspective. The study discusses how the adoption of human judgment and feedback can address critical gaps in accuracy, accountability, fairness, transparency, and trust, especially when combined with explainable AI (XAI) techniques. However, such a human oversight paradigm can also raise issues, including scalability constraints, human error, cognitive load and fatigue, as well as biases in training data, privacy risks, “rubber stamping”, or even legal liability. These concepts, advantages and challenges are are first analysed through a literature review, and then further explored through the AI-DAPT project, highlighting the implementation of techniques and measures for embedding human oversight in adaptive AI-Ops pipelines. The paper further examines the AI-DAPT healthcare domain and its pilot case and reflects on technical, ethical, and legal insights, including the future role of human oversight in the context of Agentic AI.