AI regulatory sandbox rules take effect December 1, and the fine print finally says what the sales pitch would not: a company can test its system with a regulator watching, walk out with an official record of the exercise, and still owe every euro of damage that system does to a human being. Brussels did not build a safety valve. It built a receipt printer.

The European Commission adopted the common rules on November 10, published Implementing Regulation (EU) 2026/2487 in the Official Journal the next day, and set the start date at December 1. From then on every national sandbox runs to one template: a plan agreed in advance, an authority supervising the work, written proof of the activities, and an exit report delivered no later than 45 working days after a project ends. Tech Times EU AI Act Service Desk

Those two documents have a peculiar legal status. Market surveillance authorities and notified bodies in every member state must recognise them and are told to weigh them positively when the company eventually goes through conformity assessment. They are explicitly not a declaration of conformity, not a shortcut to market access, and not a verdict on whether the system is safe. Which means the participant’s reward for cooperating with supervision is a well-formatted note confirming that it cooperated with supervision.

AI Regulatory Sandbox — The Only Thing It Clears Is the Regulator’s Calendar

Strip the language down and the bargain looks like this. Participation is voluntary. The provider stays fully responsible for complying with EU law. Liability for harm to third parties does not move an inch. The one real benefit is conditional: administrative fines for AI Act infringements are barred, but only while the company observes the agreed plan and follows the authority’s guidance in good faith, and the plan can be revised only with the authority’s written agreement, so improvisation is not a defence. Protection under other laws arrives only where those other regulators were actively in the room. Superpower Daily

Now read that from the position of the person on the receiving end. The rejected job applicant. The claimant whose file got scored. The patient whose scan went to a triage system. Nothing in the regulation improves their position by a single day. The system still ships, the record that it was tested stays largely out of their hands, and the company that hurt them has already been through a supervised exercise designed to make the eventual lawsuit cheaper to defend.

Deirdre Ahern — The Trap Was Visible in the Fine Print Before the Print Existed

Deirdre Ahern, a law professor at Trinity College Dublin, published a peer-reviewed analysis months before the Commission’s text landed, and predicted the hole the adopted regulation now leaves open. Her argument: if a sandbox records carry no conformity status, a provider with a competent compliance team has little reason to queue for one, because the ordinary standards route can deliver a presumption of conformity that no exit report can. Sandboxes, on that reading, end up attracting the firms with no cheaper path rather than the ones with the most to test. arXiv

That is the selection effect missing from the announcement. Free core guidance and supervision for small firms with an EU registered office or branch is a genuinely useful offer, and also a magnet for exactly the projects nobody has stress-tested. Non-EU providers can apply too, but only alongside an authorised representative inside the Union, and only paying the fee. The queue will be shaped by who cannot afford any other door.

Border Control — Europe Wrote Itself an Exemption and Called It Confidentiality

The transparency terms are the part worth reading twice. An exit report reaches the EU platform only when the company and the authority both agree to publish it. Without that consent an authority may still release key findings and lessons, but only non-confidential material, and information counts as disclosable unless it was declared confidential in the sandbox plan or flagged by the participant before the project ended.

Then comes the carve-out. Operational data from law enforcement, border control, immigration and asylum sits outside those disclosures entirely. So the systems with the sharpest human consequences, the risk score attached to an asylum claim, the watchlist match at a border gate, the patrol map that decides which street gets watched, are precisely the ones whose testing record Europe has pre-agreed to keep quiet. A person whose interview was scored by a system that went through a sandbox will not read the exit report, because the rulebook has already decided that report was never for her.

The wall is glass. It faces the company.

Member States — A Sandbox Network Due a Year After the Rules Already Bit

The timetable indicts itself. The Act’s high-risk obligations began applying in August 2026. Each member state now has until August 2, 2027 to have at least one sandbox operational. That leaves a stretch of roughly twelve months in which the supervision Europe designed for its most consequential systems exists as a December rulebook and a press release, while companies are already deploying under the obligations themselves. A regulator that cannot convene a sandbox is not supervising. It is corresponding.

The exits are soft by design as well. Authorities may suspend testing, temporarily or permanently, when significant risks to health, safety or fundamental rights cannot be effectively reduced, and they must tell the Commission’s AI Office why without undue delay. Extensions are available on request up to 30 calendar days before a project’s planned end. What the rules never promise is that a suspended project, or a failure inside one, becomes a public document. Serious incidents get logged in the exit report that may never be published.

Axel’s Take — Supervision Without a Published Record Is a Subscription

Here is what the arrangement buys a European taxpayer. A company receives free or subsidised supervision, a defensible paper trail, and conditional immunity from the Act’s administrative fines. The public receives annual reports from each national authority and a platform that fills up only when the regulator and the company agree to fill it. The distribution of benefit is not subtle, and neither is the incentive it creates: cooperate with the regulator, keep the receipts, tell the harmed party nothing.

The repair does not require reopening the Act. Publish every exit report by default, with a narrow, time-limited redaction for genuine operational security detail rather than a blanket exclusion covering four of the most consequential use cases in the regulation. Make the fine shield contingent on publication, so the immunity is earned in public or not at all. And tie third-party liability to the record, so a company that wants the credibility of supervised testing also carries the cost when the tested system mangles somebody’s life. This is what AI-mekna looks like when the regulator’s answer to a moving target is documentation.

Europe spent three years telling the world it would be the first to make AI answerable. What it has produced, so far, is a sandbox in which the sand is optional, the box is voluntary, and the only party guaranteed to come out clean is the one holding the clipboard.

Sources: Tech Times, Superpower Daily, EU AI Act Service Desk, arXiv.