7JUN

EU sends AI labs loss-of-control queries

3 min read

10:36UTC

Henna Virkkunen convened the 60-member AI Scientific Panel on 9 October; with the AI Office it drafted questions for frontier AI developers about recent loss-of-control incidents.

← Back to Lightning halts Saudi warm-up for two hours Jump to analysis ↓

SportDeveloping

Key takeaway

Brussels is testing whether its AI rulebook reaches frontier labs based outside Europe.

The European Commission will send requests for information to the developers of frontier AI models, the most capable systems behind the big chatbots, about what it calls recent loss-of-control incidents⁠¹. Executive Vice-President Henna Virkkunen convened a special meeting of the AI Scientific Panel on 9 October to set the requests up. No Commission text names a company or a model.

The 60-member panel drafted the questions jointly with the AI Office, the Commission unit that supervises general-purpose AI models under the EU AI Act. The AI Office says it has already requested and received serious-incident reports from providers, and sent separate requests on their security practices⁠². the Commission's 12 October account lists the chair's four recommendations: governance of internal models during development and test deployment; a stronger EU capability to detect, evaluate and investigate incidents; an international initiative to share information and respond jointly to crises; and closer scrutiny of training techniques that make models harder to monitor⁠³.

Nearly every frontier lab sits outside the EU, so the requests test whether Europe's rulebook reaches the companies it buys its most capable models from. Until this month the Commission's general-purpose AI supervision had shown up as dates and registers: a fining power that took effect against a stale contact list, then the AI Omnibus widening the AI Office's reach to models inside large platforms. The 9 October meeting turns that machinery on incidents for the first time on the public record.

A request for information carries no penalty. Under the AI Act a fine on a provider of a systemic-risk model follows only a finding of breach, and the requests sit a step before that. Virkkunen's claim that "the EU has the first law in the world that addresses systemic risk from AI" will be judged on what the AI Office does with the answers⁠⁴.

Deep Analysis

In plain English

Frontier AI models are the most powerful AI systems, the ones behind the big chatbots. A loss-of-control incident is a case where such a system does something its makers did not intend and could not easily stop. The European Commission is the EU's executive arm. Its AI Office enforces the EU AI Act, the bloc's main AI law, for these large models. The AI Scientific Panel is a group of 60 independent experts that advises it. The panel and the office have written questions for the companies involved. Most of the world's frontier labs are based outside the EU. A request for information is not a fine; it is the step a regulator takes to gather facts before deciding whether a law was broken.

Deep Analysis

Root Causes

The AI Act puts general-purpose model supervision in one Brussels unit rather than 27 national regulators. That design lets the Commission question a lab directly, which no national authority could do for a provider with no EU establishment.

The chair's recommendation on internal models points at a gap in the law itself: the Act regulates models placed on the market, while the systems a lab tests before release sit largely outside it. Loss-of-control incidents in test deployment fall into that gap.

The fourth recommendation, on training techniques that make models harder to monitor, reflects a technical constraint: if a model can mask its own behaviour, an incident report filed by the provider is the only window the regulator has.

What could happen next?

Precedent
The AI Office now has a working route to question non-EU labs about safety incidents, which later requests can follow.
Medium term · Assessed
Risk
US trade officials may cite the requests as a further EU digital trade barrier in the 2027 National Trade Estimate.
Short term · Suggested
Consequence
Labs that answer will create the first EU record of loss-of-control incidents, which any future enforcement case would draw on.
Medium term · Assessed

Source Landscape

This story draws on neutral-leaning sources from Belgium

Belgium×2

Primary parallel: On 12 October 2023 the Commission sent its first Digital Services Act request for information to X over content linked to the Hamas attack, and opened formal proceedings that December. The request was the opening move in an enforcement file, not a stand-alone inquiry.

Counter-parallel: GDPR applied from May 2018, but the largest fine against a US platform, Meta's €1.2bn for data transfers, came only in May 2023. Information powers moved fast; penalties under a regime run through a lead authority took five years.

The fine ceiling for a general-purpose model provider is €15m or 3% of global annual turnover, whichever is higher. For the largest labs that ceiling runs into hundreds of millions of euros, which is why the request stage matters: it builds the evidence file any later penalty would rest on.

The structural cost falls on compliance teams at the labs, which must now answer the same questions in Brussels that they answer in London and Washington, under three different disclosure regimes.

Consensus view: the Commission's own account treats the requests as an extension of supervision it already runs: the AI Office says it has collected serious-incident reports and asked providers about their security practices, and the panel chair's four recommendations aim at internal models, detection capacity, crisis coordination and training methods that resist monitoring.

Counter-view: Washington reads EU digital supervision of American firms as a trade question, not a safety one. The US Trade Representative named the EU among markets to be assessed for digital trade barriers on 14 September, and Jamieson Greer said in July that the Google fine created massive uncertainty for US exports.

Key tension: Whether a request for information can extract internal-model detail from labs based outside the EU before any breach finding gives the AI Office leverage to compel it.

Sources:European Commission·European Commission

Mentions:European Commission →AI Office →Henna Virkkunen →EU AI Act →Jamieson Greer →Google →Washington →Meta →Hamas →GDPR →AI Omnibus →Digital Services Act →GPAI →

First Reported In

Update #17 · EU questions AI labs over loss of control

European Commission· 22 Oct 2026

Read original →

Causes and effects

Caused by

AI Act fines land on a stale register

The AI Act fining power that began on 2 August stands behind the AI Office's first requests on loss-of-control incidents.

Occurred 2 Aug 2026

Read story →

Omnibus widens the AI Office's reach

The AI Omnibus widened the AI Office's supervisory reach before it turned that reach on frontier incidents.

Occurred 27 Jul 2026

Read story →

This Event

EU sends AI labs loss-of-control queries

The requests are the first public test of whether the AI Act's systemic-risk rules reach frontier labs based almost entirely outside the EU.

Different Perspectives

Council on Foreign Relations

CFR's analysis identified 18.75% of the 48-team field facing access barriers — the only available quantification after the State Department admitted it has none. The explicit contrast with South Africa 2010's visa waiver for all ticket holders positions the US as a regressive host by historical standards.

New York commuters / Assemblyman Bhalla

Bhalla framed Penn Station's closures as a class issue: 'For the 99 percent that can't afford to attend the games, it would be nice to at least know how you're going to get home from work.' Eight closures lock 650,000 daily commuters out for tournament traffic, with one on a Monday rush hour.

French Football Federation

The FFF called an 11:00 CEST press conference for 28 July, at its Paris headquarters, following an extraordinary executive committee meeting, at which Zinedine Zidane is expected to be presented as head coach succeeding Didier Deschamps. As of this writing no communique confirms the appointment, contract length or start date.

Morocco

Morocco is pushing for the 2030 final at its Grand Stade Hassan II in Casablanca, a planned 115,000-seat venue that would be the world's largest football stadium on completion. The venue is contested by a Spanish petition for the Santiago Bernabeu, and FIFA has fixed neither the venue nor the tournament's proposed 64-team format.

FIFA

FIFA's first tournament report, published 27 July, itemises nearly 300,000 accredited personnel, 73,700 security staff and 300,000 square metres of custom-grown turf, with no revenue or cost figure attached. As a Swiss association answering to its own Congress rather than a treasury, FIFA has followed the same operational-before-financial sequence it used after Qatar 2022.

Javier Tebas / La Liga

La Liga president Javier Tebas said on 21 July that FIFA's system is rotten from the root and that Gianni Infantino's time as president has, in his words, concluded. Tebas has no vote in FIFA's process, so days after the IOC declined jurisdiction over a separate ethics complaint, he used the only instrument he has.

Written with AI from Lowdown's published reporting. Not individually reviewed by an editor. Editorial standards.