Designing human oversight: what does Article 14 of the AI Act require?

Short answer: Article 14 of the AI Act requires providers of high-risk AI to design their system so that a human can effectively oversee and correct it during use. Oversight must actually work: the overseer must understand the output, know when to rely on it, be able to intervene or stop, and not blindly follow the machine (automation bias). It is a design requirement, not a tick-box after the fact.

What human oversight must enable

Article 14 requires the assigned individuals to be able to:

• understand the capabilities and limitations of the system and monitor its operation, including signs of anomalies or malfunction;
• recognise and counter automation bias — the tendency to automatically trust AI output, especially in systems that generate recommendations about people;
• interpret the output correctly, with awareness of the available tools and methods;
• decide not to use the output, or to disregard, override or reverse it;
• intervene or stop the system through a stop button or comparable procedure that brings the system to a safe state.

Design, not just procedure

The core of Article 14 is that oversight measures are built in by the provider — technically in the system itself, or as measures the deployer must implement. Instructions saying "check the result" are not enough if the system does not facilitate that check. Measures must be commensurate with the risks, the degree of autonomy and the context of use.

Oversight therefore connects to other high-risk requirements: without intelligible instructions for use and transparency a human cannot interpret the output, and without record-keeping cannot reconstruct afterwards what happened.

The four-eyes principle for some biometrics

For certain remote biometric identification systems Article 14 sets a stricter rule: no action may be taken on the basis of the identification unless at least two competent persons have separately verified it. This "four-eyes" principle does not apply in specific exempted cases for law enforcement, migration or border control.

What to do

• Assign the oversight role concretely to trained, authorised individuals — not "the team" in general.
• Build a working intervention and stop function that safely halts the system, and test it.
• Surface the system's limits in the interface: confidence scores, warnings, uncertainty signals.
• Train against automation bias: have overseers practise failure situations, not just the happy path.
• Document the measures in the technical documentation and the instructions for use.
• Link oversight to governance through your AI governance framework and the timeline of obligations.

Human oversight is one of the core requirements in the overview of high-risk obligations and weighs into the conformity assessment and CE marking. A stop button no one can operate does not count.