iSearch

What is human oversight of AI?

Human oversight means that a person can follow, review and when needed stop or change what an AI system does.

It is often called having a human in the loop. AI produces a suggestion and a person decides what happens with it.

The EU AI Act requires high-risk systems to be designed so that people can oversee them. For other systems it is good practice, especially when the answers affect decisions.

In practice it is about answers being checkable. A draft that shows its sources is easy to review. An answer without a source has to be taken on trust.

How it works

Oversight can be built in at different levels. A person can approve every suggestion before it is used, review a sample afterwards or be alerted when the system is uncertain. What fits depends on how serious a mistake would be and how often it happens.

Why it matters

Human oversight keeps responsibility with a person who can stand behind the decision. It catches mistakes before they have consequences and signals where the system needs improving. It also makes it easier to show others that AI is used responsibly.

What to keep in mind

There is a risk that review becomes routine when AI is usually right. The reviewer starts trusting the suggestion and stops checking. This is sometimes called automation bias.

So make reviewing easy. Show sources, flag uncertainty and track how often suggestions are actually changed.

What it means for you

Decide which AI answers may be used directly and which must always be reviewed by a person first.

Common questions

When is human oversight required?

The AI Act requires it for high-risk AI. For other systems it is good practice, especially when answers feed into decisions about people.

Is it enough for someone to read the answer?

Only if that person has the time, knowledge and authority to change or stop it. Otherwise the review becomes a formality.

How do you make oversight practical?

Let answers show their sources so review is quick. Decide in advance which answers must always be checked.

Sources: EU AI Act, Article 14, on EUR-Lex

All terms Updated