Human Oversight under the EU AI Act: Who, What and How?

0

On 16 November 2026, MediaLaws hosted a Dialogue dedicated to one of the central questions in the implementation of the EU AI Act: what does meaningful human oversight require in practice?

The discussion focused on Articles 14 and 26 of the AI Act, examining how oversight requirements can be translated into technical design choices, organisational responsibilities and verifiable safeguards. The four speakers addressed the question from complementary perspectives, including harmonised standards, organisational governance, fundamental rights protection and agentic AI.

Luca Nannini examined the ongoing development of draft European harmonised standard EN 18229-5 on human oversight. His presentation addressed the relationship between the design obligations imposed on providers under Article 14 and the responsibilities of deployers under Article 26. Particular attention was given to the competencies required of human overseers, the time available for intervention, the verification of oversight measures and the implementation of the two-person verification requirement for remote biometric identification under Article 14(5).

Ilaria Penco presented an organisational methodology developed through research at Intesa Sanpaolo, distinguishing operational users, who interact with individual AI outputs, from human overseers responsible for monitoring a system’s functioning. Meaningful oversight, in this account, depends on competence, critical independence, authority and effective intervention powers. The discussion also considered the importance of ex post monitoring for identifying systemic risks that may remain invisible at the level of individual decisions.

Irina Carnat addressed human oversight as a sociotechnical safeguard, examining the allocation of responsibilities between providers and deployers and the risk that human intervention becomes a merely formal exercise. Drawing on the GDPR framework and the SCHUFA judgment, her contribution explored automation bias, the limits of explainability techniques and the conditions necessary for human reviewers to exercise independent judgment rather than simply endorse algorithmic outputs.

Abdullah Elbi considered the implications of agentic AI for existing oversight models. Increasingly autonomous systems raise questions about when human intervention remains technically feasible and how oversight requirements can be reconciled with cybersecurity considerations. His presentation examined the tensions between treating humans as safeguards against AI risks and recognising that human intervention may itself introduce vulnerabilities.

A recurring question throughout the discussion concerned the effectiveness of oversight beyond individual outputs. Human participation does not necessarily provide meaningful control where overseers lack sufficient information, independence, time or authority to intervene. Equally, risks affecting fundamental rights may accumulate across decisions and become apparent only through monitoring at system or organisational level.

The discussion therefore examined human oversight as a continuing governance function, whose effectiveness depends on both the technical measures established by providers and the organisational arrangements implemented by deployers. It also raised questions about how these arrangements should be verified, documented and coordinated with other obligations under the AI Act.

Watch the full webinar recording: https://www.youtube.com/watch?v=YmknutMNfXw

Share this article!

About Author

Redazione MediaLaws

Leave A Reply