Article 14 and the Problem of Oversight Theatre

Article 14 of the EU AI Act sounds almost trivial: high-risk AI must stay under effective human oversight. The word doing the work is effective, and it is where most implementations quietly fail.

The failure has a name: oversight theatre. A person is placed in the loop, an approval button is added, a sign-off is recorded, and on paper there is human oversight. In practice the person cannot question the system in any meaningful way, lacks the information or the time to intervene, and approves by default. That is oversight in form, not in substance, and it does not satisfy Article 14.

Read closely, the article asks for more than a human presence. It asks that the people overseeing a system can understand its capabilities and limits, stay alert to automation bias, interpret the output correctly, decide not to use it, and intervene or stop it. Each of those is a design requirement, not a policy statement, and a system that makes any of them impossible has an oversight defect.

This is why oversight is a technical problem before it is an organisational one. The interface has to surface the right information at the right moment, expose uncertainty rather than hide it, and make intervention genuinely possible within the time the decision allows. Oversight bolted on after the model is built tends to be the nominal kind.

There is a direct line from Article 14 to Article 4. Oversight by a person without sufficient literacy is oversight in name only, so the literacy obligation and the oversight obligation are two halves of the same requirement. Competent oversight assumes a competent overseer.

The hard balance is utility. Oversight that forces a human to re-derive every output destroys the value of the system, while oversight that asks nothing is theatre. The design target is the middle: meaningful checkpoints at the moments that matter, calibrated to consequence, leaving the routine to run.

Documenting all of this belongs in Annex IV, where the oversight measures are recorded as part of the technical file, and the documentation should describe oversight that actually works, not oversight that was assumed.

Our whitepaper, Designing for Human Oversight, reads Article 14 in full, sets out the core obligations and their technical implementation, connects oversight to literacy, and shows how to avoid theatre while keeping the system useful. Effective oversight is a design discipline, and it is one worth getting right.