Article 14: human oversight for high-risk AI

Article 14 says high-risk AI systems must be built so that people can effectively oversee them while they are in use. The point is simple: a system that affects people's lives cannot be left to run on its own.

Oversight has to be real, not a box to tick. The people overseeing a system need to be able to understand it, watch how it is performing, and step in when something is wrong.

An overseer should be able to understand the system's capabilities and limits, and stay alert to signs it is not working as expected. They should be able to interpret the output correctly, and know when to disregard, override, or reverse it.

They also need a way to act: to stop the system, or to decline to act on its output. For some systems the Act expects a clear way to halt operation, sometimes described as a stop button.

Article 14 is mainly a duty on the provider, the company that builds the system, to design in the means for oversight. But it reaches you as a deployer through Article 26.

Article 26 requires you to assign human oversight to natural persons who have the competence, training, and authority to do it, and to give them the support they need. Buying a system with oversight features is not enough; you have to put real people in the loop and back them.

Article 14 names a specific risk: automation bias, the tendency to over-trust a system's output simply because a machine produced it. A reviewer who rubber-stamps every result is not providing oversight.

Effective oversight means the overseer is genuinely able, and expected, to disagree with the system. That takes time, the right information, and a culture where overriding the tool is normal rather than career-limiting.

Name actual people, not a job title nobody owns. They need enough understanding of the system to catch its mistakes, the authority to act on what they find, and the time to do it properly.

For sensitive uses, some independence helps. The person checking a system's decisions should not be the same person under pressure to hit the numbers the system is meant to improve.

In recruitment, oversight might mean a person reviews the shortlist a system produces, can see why a candidate was ranked low, and can add candidates back in. In lending, it might mean a person can review and reverse an automated decline before it is sent.

The common thread is that a capable human can see what the system did, judge whether it is right, and change the outcome.

Write down how oversight works for each high-risk system: who is responsible, what they watch for, when they intervene, the escalation path, and how their oversight is recorded. That record is what you show an auditor.

Veillo generates a Human Oversight Protocol per high-risk system as a structured starting point, which you adapt to name the real people and thresholds.