Section 14 of the AI Act: How to Demonstrate Human Oversight

Approving is not the same as supervising.

November 16, 2026Reading time: 7 min
Short answer

Article 14 of Regulation (EU) 2024/1689 requires that high-risk AI systems be monitored by people who understand what the system does, can detect when they are relying too heavily on it, know how to interpret its output, and can override or shut it down. A log entry that simply says “approved” is not sufficient proof of this. It must be possible to reconstruct every decision: who reviewed it, what they saw, when, and whether they could actually have said no.

Think of a typical approval workflow: a green button, a name, and a date. Now ask yourself what that person saw before clicking. If the answer is “we don’t know,” that’s the gap this article aims to fill.

What Article 14 Says, in Plain Language

  1. 14.1The high-risk system must be designed so that individuals can effectively monitor it while it is in use, including through the use of appropriate interface tools.
  2. 14.2The objective is to prevent or reduce risks to health, safety, or fundamental rights, even when the system is used in a foreseeable manner that constitutes misuse.
  3. 14.3Measures must be proportionate to the risk, the system’s autonomy, and the context. They are implemented by the provider or applied by the party deploying the system.
  4. 14.4The person supervising must be able to: a) understand the system’s capabilities and limitations and detect anomalies; b) be aware of automation bias—the temptation to trust the system without thinking; c) interpret the result correctly; d) decide not to use it, ignore it, override it, or reverse it; e) intervene or stop the system using a stop button or something equivalent.
  5. 14.5In remote biometric identification, no decision is made without being independently verified by at least two competent individuals.

Who Does What

Supplier
Design the brake

Design the system so that it can be monitored (Art. 14) and keeps automatic records of its operation (Art. 12).

Party Responsible for Deployment (Your Company)
It shows that someone was able to brake

Assign supervision to individuals who are competent, trained, and authorized (Art. 26.2), and retain the records for at least six months (Art. 26.6).

When does it apply?

Today · Spain
CGPJ, Directive 2/2026

Mandatory human oversight of AI in the justice system. See analysis

Dec. 2, 2027
High Risk · Appendix III

Employment, credit, essential services, education. What comes next after the Digital Omnibus?

Aug. 2028
High Risk · Appendix I

AI integrated into regulated products.

Companies that wait until 2027 will be too late for their own audit.

What evidence do you need, requirement by requirement?

RequirementEvidenceHow to Make It Verifiable
Understanding theSystem 14.4.aPersonnel training, documentation on usage received, and the version of the form in effect on that day.Certify the version of the form and the training record with a certified date.
Automation Bias14.4.bCancellation rate, review time, bulk approval alerts.Record those metrics where no one can change them later.
Interpreting theresult 14.4.cWhat you saw on the screen: input, result, and confidence score.Save a cryptographic hash of that context, not just the “OK.”
Do not use it ordisable it 14.4.dThe decision, the reason behind it, and who made it.Signature and timestamp.
Intervene orstop 14.4.eThat the bus stop existed and was in use, and every time it was used.Stamped record of periodic shutdown tests.

The Five Questions an Auditor Will Ask You

  1. 1Who oversaw this decision and had the authority to stop it?
  2. 2What information did you have available when you made your decision?
  3. 3Which version of the model was in production that day?
  4. 4How many times has your team said “no” to the system this quarter?
  5. 5Can you prove that these records have not been altered since then?

If the fifth question makes you uncomfortable, you're not alone. Most records are stored in the same system that is being audited, and the person who administers it can edit them.

Common Mistakes

Approve 100 decisions all at once with a single click. A “human in the loop” who has never overruled anything. Records kept in a shared spreadsheet. Training with no date or signature. And the most costly issue of all: no one knows who can shut down the system on a Friday afternoon.

An example: credit scoring

He saw
Score, 3 factors, confidence
He decided
Cancel the recommendation
Reason
Income Misinterpreted by the Model
When
Date and time stamped

A bank uses a model to score credit applications—a high-risk case from Annex III. The analyst reviews the score, the three most significant factors, and the confidence level. She rejects the recommendation because the system did not correctly interpret the customer’s income. That moment—what she saw, her decision, her reasoning, and the time—is what must be documented two years later. We explain this step by step in the practical case study on high-risk AI governance.

Frequently Asked Questions

Is a "pass" checkbox enough?

Not on its own. It shows that someone clicked on it, not that they were able to understand it and shut down the system.

How long do the records need to be kept?

At least six months, according to Article 26.6, unless another regulation—such as one governing the financial sector—requires a longer period.

Does this apply to generative AI?

Only if it's used in a high-risk situation. A customer service chatbot isn't one. A model that screens candidates is.

What about AI agents?

Yes, if they are working in a high-risk environment. With agents, it is even more important to document what they were able to do and who authorized each action.

Sources

Regulation (EU) 2024/1689, Articles 12, 14, and 26 (EUR-Lex) · CGPJ, Instruction 2/2026 (analysis in the Journal) · V-PROOF: Implementation Timeline for the EU AI Act and the Digital Omnibus.

Next
Next

AI-to-Human Ratio: What It Is and How to Demonstrate It in an Audit