Since May 2026 I sit inside a room where most people never go. It is called CEN-CENELEC JTC 21, Working Group 4. It is the group writing the technical standards that will tell companies what the EU AI Act actually means in practice.
The Act says high-risk AI systems must be designed so that humans can "effectively oversee" them. Article 14. Nice sentence. But what is effective? Who decides? In Europe, the answer is: a standard. For human oversight, this standard is called prEN 18229-3. When it is cited in the Official Journal, it becomes the way a provider shows compliance. So the text of this standard is, in practice, the real law.
On 25 August 2026, I filed four technical comments on it, through the public enquiry of the British Standards Institution. Later the same comments, and four more, went to CEN through Malta's national standards body. This page explains the four, in plain words, because I think the people who worry about AI going wrong should know what is being decided here, and when.
1.Oversight is a dial, not a switch Clause 5.2.6
The draft asks the provider to choose an oversight configuration at design time. How much a human checks, how often, how fast they can react. Then you review it when something unexpected happens: a design change, a problem seen in post-market monitoring.
The problem: it is always reactive. Something must go wrong first.
In real life, nobody runs an AI system like this. You start with a human checking everything. Then, when you have evidence the outputs are good, you check less. Pilot, supervised rollout, routine operation. Everybody does it. The standard has no word for it.
So today a provider has two bad options. Declare the most strict oversight forever, which is not true in practice. Or treat every reduction as a "design modification", with all the paperwork, which punishes the provider who planned properly.
Let the provider declare the whole sequence in advance. Each step, the evidence needed to move to the next one, the minimum time and data before moving, and — the most important part — the conditions to go back. And every step must still respect the reaction time the risk needs.
Why I care: I run a learning platform. Oversight on it is heavy at the start and lighter when we know a cohort. It works. But I also see the danger. Reducing oversight because things went well for a while is exactly how oversight erodes without anybody deciding it. Writing the reversion conditions down, before you need them, is the protection.
2.Where does the stop button go? Clause 5.2.5
The draft thinks about one person overseeing several AI systems. Good. It does not think about one system made of several AI components, one feeding the next.
This is more and more how systems are built. A model plans, another one retrieves, another one writes, another one checks. If the human can only stop the final output, the cause upstream is untouched; next time it comes back. If the human must look at every intermediate output, they are overloaded in one afternoon, and then they stop really looking.
The provider must say where in the chain the human can intervene, and justify why that point is enough. Not everywhere. Not only at the end. At the boundary that actually controls the risk, with a written reason.
3.The overseer is a human Clause 5.5.1
The draft talks about automation bias — the tendency to trust what the machine says. Good again. But it imagines the overseer as a trained professional.
In many high-risk uses, the overseer is a teacher, a nurse, an HR assistant, sometimes the end user. Not a specialist. They have little time and a lot of trust in the screen. My comment asks the standard to treat that person as the normal case, not the exception, and to say so explicitly where it handles automation bias.
4.Small providers are not small risks Clause 5.3.1
The verification work the standard expects is designed with large organisations in mind. For a small company the cost is real. My comment asks for proportionate ways to meet the same objective — not a lower bar, a path a five-person team can actually walk. Otherwise small providers will tick boxes, and box-ticking is worse oversight than no claim of oversight at all.
Why this is bigger than hiring algorithms
To be honest about limits first. The high-risk part of the AI Act covers things like CV screening, credit scoring, education, medical devices. It is not about frontier models or loss of control. Writing good text in prEN 18229-3 will not stop a catastrophe by itself.
But look at what "keep a human in the loop" is. It is the first answer everybody gives, for a hiring tool and for much more powerful systems. And this standard is the first place in the world where "effective human oversight" gets a precise, binding, auditable meaning. Whatever we write here becomes the reference people copy later.
And the three weak points I found are the same at every scale:
- oversight that quietly decreases because things went well;
- oversight placed at the end of a chain, where it sees the result but not the cause;
- overseers who are not equipped to doubt the machine.
For a learning platform these are compliance problems. For stronger, more autonomous systems, they are exactly the ways humans lose control without noticing. Better to write the protections now, in the boring document, while the stakes are still small.
The uncomfortable part
I prepared these comments, and this page, with an AI assistant. It helped me read, structure, translate my French-English into something a committee can use. The arguments come from my platform and from my work in the group; the assistant helped me say them.
Which is exactly my third point. If the human who oversees is also helped, shaped, maybe a little bit steered by the system — how independent is the oversight? I do not have the full answer. I think nobody has it yet. But I would prefer we ask the question inside the standard rather than discover it after.
- prEN 18229-3 — Human oversight Public enquiry closed on 22 September 2026. Comments are now in resolution. I do not know yet which ones will be accepted. I will update this page.
- prEN 18229-2 — Transparency Next in line. When its enquiry opens, I will publish the dates here, and how anybody can comment through their national standards body, or through the BSI public portal, which is open to everyone.
- If you work on AI safety and you never read a harmonised standard, this is a good moment to start. The texts are being frozen in the next twelve months. The obligations apply from December 2027.
You can find me on LinkedIn: linkedin.com/in/victorlarde369. Questions, disagreement, all welcome — or write to info@unchainacademy.com.