On the sixth of May, with less than three months to go before the EU AI Act's high-risk obligations were due to bite, European lawmakers agreed to push the deadline. Education systems sit in Annex III of the Act, the high-risk list, alongside hiring, credit, and border control, and providers of those systems were preparing, with varying sincerity, for August 2026. The omnibus agreement moves standalone Annex III obligations to December 2027.
I build an adaptive learning platform, so this is my category, and I have complicated feelings about the extra sixteen months that I want to work through in public.
What the delay actually says
The official reasoning is readiness, the standards bodies are behind, the guidance is unfinished, and enforcement against rules nobody can operationalize helps no one. All true as far as I can tell. And underneath the procedural truth sits a quieter one. The companies that asked for relief got it, the families using these products were not at the table, and most of them never knew a deadline existed in the first place. A regulation written because AI systems were already shaping children's education has now twice agreed to wait while those systems keep shipping.
I do not say that with outrage. I say it as a plain description of how the incentives sit, because the practical consequence lands on builders like me. Until the rules arrive at the end of 2027, conformity in education AI is mostly voluntary, and voluntary conformity is worth exactly as much as the engineering behind it.
What the rules would actually ask of me
Here is the part I find most useful to write down, because the obligations are less mysterious than the compliance industry makes them sound. Reading the deployer obligations and the provider duties they mirror, and translating them into the shape of a learning product, the law essentially asks five things.
It asks that human oversight be real, a person able to understand the system's capabilities and limits, intervene, and overrule it. On my platform that is the review pipeline and the guardian controls, a person in the chain from generation to serving, with the authority to stop anything. I wrote last summer about where that authority sits and what the machine may clear on its own.
It asks that outputs be interpretable to the person overseeing them. A reviewer who cannot tell why a problem was approved cannot oversee, only rubber-stamp. This is why my review queue shows the scores, the gate results, and the rendered screen together, although I would not claim that problem fully solved.
It asks for logging, the record of what the system did and when, kept long enough for someone to reconstruct an incident. It asks for risk management as a practice, not a binder, which in my world means the failure modes of generated content are enumerated, tested, and re-tested as the system changes. And it asks that the people affected know what the system is and is not doing, which for a children's product means the guardians, in language families actually read.
Stated that way, the list is not regulatory exotica. It is close to what I think building seriously for children requires anyway. That is precisely why the delay sits uneasily with me, the builders for whom these obligations are a paperwork exercise just got sixteen months of slack, and the builders for whom they were already engineering practice did not need the slack in the first place.
The admission that keeps this post honest
It would be easy to end there, with the well-behaved builder politely disappointed in his lazier neighbors. So let me put the uncomfortable part on the record. Voluntary compliance is also easy to claim and hard to verify, including mine. Everything I have described, the gates, the thresholds, the oversight, you have my word for, and my word is exactly the standard of evidence the delay leaves the whole category running on. The thing regulation actually adds is not virtue, it is verification, an external party with the power to ask for the logs. Absent that, most safety claims in education AI, this blog included, are marketing until somebody audits them.
That is the real cost of December 2027. Not that builders will suddenly behave worse, most will not, but that families have no way to tell the difference between engineering and theater until the end of 2027, in a category where they could not tell anyway and the users cannot advocate for themselves.
What I think builders should do with the time
If you build in this category, my suggestion is unoriginal and I will make it anyway. Build to the standard now and write down what you did, dates and all. Run the conformity exercise against your own system this year, while it is cheap, while changing the architecture is a refactor instead of a recall. The documentation that feels bureaucratic today is the receipt that proves your safety story was not retrofitted the week before enforcement.
And if you evaluate these products, as a parent, a school, a district, the delay hands you a useful question that costs nothing to ask. Tell me what you do today, voluntarily, that the AI Act would have required of you in August. The builders with a real answer tend to be specific and a little tired about it. The ones without one often talk about their values instead, and you can usually hear the difference quickly.
I might be reading the moment wrong. Maybe the standards arrive mature in 2027, enforcement lands well, and the delay turns into the boring administrative footnote its defenders promise. I would happily be wrong that way. But rules postponed have a way of being postponed again, and children's products have a long history of being regulated one scandal too late. Sixteen months is a long time in this field. It is also two school years, and nobody enrolled in those years gets them back.