Glossary · EU AI Act: high-risk AI
Human oversight (AI Act)
Also known as: human oversight, human-in-the-loop, human review
German: Menschliche Aufsicht
Under the EU AI Act, human oversight is the requirement that a high-risk AI system be designed so that natural persons can effectively oversee it while it is in use. The people assigned to oversight must be able to understand its capacities and limitations, interpret its output, decide not to use it, override or reverse its output, and interrupt it.
- AI regulation
- EU
In one sentence
Human oversight under Art. 14 AI Act means people can understand, override and stop a high-risk AI system, with defined authority and competence.
Example
An AI system that flags defective welds on a car body line lets the quality inspector reject or confirm each flag, shows the confidence and the image region used, and has a stop function that returns the line to manual inspection.
How it applies
- Provider and deployer: The provider builds oversight measures into the system or specifies them for the deployer to implement. The deployer must assign oversight to people with the necessary competence, training and authority.
- What oversight must enable: Understanding the system's capacities and limitations, staying aware of automation bias, interpreting output correctly, deciding not to use or to disregard, override or reverse the output, and intervening or stopping the system in a safe state. For remote biometric identification, a decision based on the output must be verified by at least two people, with exceptions.
- Technical documentation: Oversight measures appear in the technical documentation and must be explained in the instructions for use, including the tools that help interpret output.
- Machine safety: In machinery, a stop function for AI does not replace safety functions designed to ISO 13849-1 or IEC 62061.
Compared with the USA, Canada and China
The USA has no general federal oversight duty; the NIST AI RMF treats human oversight as part of governance, and Colorado's replacement law requires meaningful human review of adverse consequential decisions from January 1, 2027. Canada's Directive on Automated Decision-Making requires that final decisions at the higher impact levels be made by a person, and the voluntary code includes human oversight and monitoring. China's rules make service providers responsible for reviewing and moderating content rather than prescribing oversight of individual decisions.
Read more on AI TechDoc Press
The in-depth analysis behind this entry, in the AI TechDoc Press newsletter. Subscribe for free