Module 3 — Human Oversight That Survives a Review

Module 3 of the AI Act Deployer certification: Art. 14 as a design duty on the provider, Art. 26(2) as an assignment duty on the deployer, automation bias, and how to evidence oversight that is real.

Human oversight is the obligation most often claimed and least often evidenced. Organisations write "a human reviews all decisions" into a policy and consider the matter closed. A supervisor reading that sentence will ask three questions the policy does not answer: which human, with what authority, and how would you know if they had stopped actually looking.

This module is about answering those three.

Two duties, two parties

Art. 14 binds the provider. High-risk systems shall be designed and developed, including with appropriate human-machine interface tools, such that they can be effectively overseen by natural persons during the period in which they are in use.

Art. 26(2) binds the deployer. It must assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support.

Neither discharges the other. A perfectly designed system with an untrained, unauthorised overseer fails. A well-trained, empowered overseer using a system that exposes no controls also fails — and that failure is the provider's, which makes it a procurement question for you.

What Art. 14 actually asks the system to permit

Art. 14(4) lists what the oversight person must be enabled to do. Read it as a checklist against any system you are buying:

• Understand the capacities and limitations of the system and monitor its operation, so that signs of anomalies, dysfunction and unexpected performance can be detected and addressed as soon as possible.
• Remain aware of automation bias — the tendency to automatically rely on or over-rely on output, particularly for systems used to provide information or recommendations for decisions to be taken by natural persons.
• Correctly interpret the output, taking into account the interpretation tools and methods available.
• Decide, in any particular situation, not to use the system or otherwise disregard, override or reverse the output.
• Intervene in the operation or interrupt the system through a stop button or a similar procedure that allows it to come to a halt in a safe state.

The fourth and fifth are the ones to test in a demonstration. Ask the vendor to show you the override. If the answer is that the decision can be changed downstream in another system, the override does not exist in the sense Art. 14 means.

Why automation bias is named in the text

It is unusual for a regulation to name a cognitive failure mode. Art. 14(4)(b) does, because automation bias is the specific mechanism by which oversight becomes theatre while everyone involved believes it is functioning.

The conditions that produce it are well understood and entirely ordinary: high volume, time pressure, an output presented with apparent confidence and no uncertainty, an interface where agreeing takes one click and disagreeing takes a form, and a working culture where overriding the model requires justification but following it does not.

Every one of those is a design and management choice, which is why the Regulation treats awareness of the bias as an oversight requirement rather than a footnote.

Practical counter-measures that also produce evidence:

• Show uncertainty, not just a verdict. An overseer who can see that the model is at 51% behaves differently from one shown a bare recommendation.
• Make override symmetrical. If agreeing is one click, disagreeing should be one click plus a reason — not five screens.
• Measure the override rate. A rate that is stable and non-zero is the single best evidence that oversight is live. A rate of zero over months means either a perfect model or an inert overseer, and it is never the first.
• Rotate and sample. Overseers reviewing a random sample with feedback stay calibrated; overseers reviewing everything stop reading.

Art. 14(5) — the four-eyes rule for biometrics

For the remote biometric identification systems in Annex III point 1, Art. 14(5) adds a specific requirement: no action or decision may be taken on the basis of the identification unless it has been separately verified and confirmed by at least two natural persons with the necessary competence, training and authority.

There is a carve-out for law enforcement, migration, border control and asylum in certain circumstances, but for commercial deployments the two-person rule stands. If you are deploying remote biometric identification and one person can act on a match, you are not compliant, however good the model is.

Turning "we have oversight" into evidence

The claim is unfalsifiable; the artefacts are not. For each high-risk system, hold:

The last one is worth doing literally. A stop procedure nobody has ever executed is a document, not a control.

The relationship with Art. 4

Art. 4, as rewritten by the Digital Omnibus in July 2026, requires providers and deployers to take measures supporting the AI literacy of staff who operate AI systems on their behalf. It does not require you to guarantee a level for any individual, and it carries no fine band of its own.

Art. 26(2) does carry one, through Art. 99(4). So the practical position is this: your general AI literacy programme answers Art. 4, and the specific, system-tied training of your named overseers answers Art. 26(2). The second is narrower, deeper and enforceable. Do not let the first stand in for it.

Check yourself

1. Our reviewer sees every case and approves nearly all of them. — Look at the override rate. A rate at or near zero over a long period is evidence of automation bias, not of a good model.
2. The vendor says the decision can be reversed later in our case management system. — That is not the Art. 14(4)(d) override. The oversight person must be able to disregard or reverse the output of the system itself.
3. We deploy face matching against a watchlist and one trained officer acts on matches. — Art. 14(5) requires two. Separate verification by two competent, trained and authorised persons before any action.
4. We run an annual AI awareness course, so Art. 26(2) is covered. — No. Art. 26(2) requires competence, training, authority and support tied to the specific system; a general course answers Art. 4, which carries no fine of its own.

Previous: Module 2 — Article 26, paragraph by paragraph Next: Module 4 — Logs, monitoring, suspension, incidents →