Human in the loop: the difference between a great decision and a disaster?
image: Steve A Johnson on Unsplash
What does it actually mean to have a “human in the loop” when an AI system makes decisions about people’s lives? And is human oversight enough to prevent things from going wrong?
These are the questions at the heart of a recent podcast episode and a newly published book chapter by Melanie Fink, Associate Professor of European Law at Leiden University’s Europa Institute.
The podcast
Fink joined Alba Ribera Martinez on The Binary Agora (episode 29) to discuss her research on human oversight of AI systems used by public authorities — from border control to tax administration. The conversation unpacks why “human-in-the-loop” is less a solution than a starting point. As Fink explains, there is no single image of what a human in the loop actually looks like: it can mean an engineer at the training stage, a civil servant reviewing individual outputs, or simply someone who can press stop. And each of these demands something very different from the person involved.
A key theme is the tension between formal responsibility and actual agency. Civil servants may be formally accountable for AI-assisted decisions while lacking the understanding or time to meaningfully influence them — a dynamic Fink describes as setting the human up for failure. Drawing on a three-part framework distinguishing output-oriented, process-oriented, and accountability-oriented roles, she argues that human oversight works best when its goals are clearly defined in advance and matched to tasks humans can actually perform well. Humans are not great at double-checking the accuracy of algorithmic outputs, for instance, but they can provide the procedural safeguards — the “ear” — that affected individuals are entitled to.
The episode also addresses the risk of over-reliance: placing a human in the loop should not become an excuse to dial down other safeguards. As Fink puts it at the close of the conversation, the human should not be the only thing standing between a great decision and a complete disaster.
The episode is available on YouTube and Spotify.
The book chapter
The podcast draws on Fink’s chapter “Human Oversight (Article 14)” published in The EU Artificial Intelligence Act: A Thematic Commentary, edited by Gianclaudio Malgieri, Gloria González Fuster, Alessandro Mantelero, and Gabriela Zanfir-Fortuna (Hart Publishing, May 2026). The chapter offers a systematic analysis of Article 14 of the EU AI Act, the provision that requires providers of high-risk AI systems to build the conditions for effective human oversight into their systems. Fink examines what “effective” oversight means in practice, organising the Act’s requirements around three markers: the authority of the human overseer, their comprehension of the system, and the environment in which they operate. The chapter concludes with an assessment of the limits of human oversight, cautioning against the assumption that putting a human in the loop on its own automatically makes decisions safer or fairer.
The book is available via Bloomsbury. Those without access are welcome to contact Melanie Fink directly for a copy of the chapter.