The trade the EU just made

On 7 May, negotiators for the European Parliament and Council reached political agreement on the Digital Omnibus on AI. The headline, in most of the coverage I've read, is regulatory relief: the high-risk obligations of the AI Act, originally due to bite in August 2026, have been pushed back to December 2027 for standalone systems and August 2028 for AI embedded in regulated products. Industry got most of what it asked for. The package also adds something industry did not particularly ask for: an explicit, EU-wide ban on non-consensual "nudification" apps, with a compliance deadline of December 2026.

The two halves of the deal are being reported as if they sit on the same moral plane. They do not. One is a concession to firms that found the compliance timeline inconvenient. The other is a belated answer to a harm that has been documented, in schools and police reports across the continent, for at least three years. Treating them as a balanced package, a give-and-take, obscures what actually happened: the people who could lobby got a delay, and the people who could not lobby, overwhelmingly girls and women whose images have been weaponised by free apps, got a ban that should have existed already.

I want to take the trade seriously on its own terms before saying what I think of it.

The case for the delay. The argument from industry, and from the Commission officials who accepted it, is not unreasonable on its face. The harmonised standards that high-risk system providers are meant to conform to have not all been published. The AI Office is still building capacity. National competent authorities, in many member states, do not yet exist as functioning bodies. Asking firms to demonstrate conformity with a regime whose technical specifications are incomplete is, the argument runs, asking them to comply with a moving target. Better to delay than to enforce a half-built rulebook badly.

I find parts of this persuasive and parts of it not. The standards delay is real. But the standards delay is also, in part, a consequence of how slowly the standardisation bodies have moved, and the standardisation bodies have moved slowly partly because the firms most affected have been well-represented in those rooms and have not been in a hurry. "The rules aren't ready" is a true statement that elides who was responsible for getting them ready.

Who gets the eighteen extra months. High-risk AI under the Act covers systems used in employment decisions, education and vocational training, access to essential public and private services, law enforcement, migration, and the administration of justice. These are not abstract categories. They are the systems that decide whether you get the interview, whether your child gets the school place, whether your benefits claim is flagged for fraud review, whether the border officer's screen lights up when you present your passport. The eighteen-month delay is not a delay on chatbots or image generators in the abstract. It is a delay on the obligations that were specifically designed to give the people on the receiving end of these systems some recourse.

What does the delay actually defer? Risk management systems. Data governance documentation. Human oversight requirements. Accuracy, robustness and cybersecurity testing. Transparency obligations to deployers. Registration in the EU database. Post-market monitoring. The ability of a national authority to request the technical documentation and actually read it.

That is the universe of systems whose accountability infrastructure has just been postponed by a year and a half. The firms deploying them get more time. The people subject to them get more time too, more time being subject to systems whose providers do not yet have to demonstrate that the systems work the way they claim.

The nudification clause. The other half of the deal is the explicit prohibition on AI applications whose primary purpose is generating non-consensual sexual imagery of identifiable people. The compliance deadline is December 2026, eight months from the political agreement, and the prohibition reaches both providers and distributors, which means app stores and hosting infrastructure as well as the developers themselves.

This is, unambiguously, a good provision. The harm it addresses is concrete and well-documented. Internet Watch Foundation reporting, school safeguarding data from Spain, France, the Netherlands and Italy, and academic work on image-based sexual abuse have established the pattern: free or near-free apps, marketed openly on mainstream platforms, used overwhelmingly against women and girls, with effects on the victims that are severe and lasting.¹ The Article 5 prohibitions in the AI Act, the existing ban list, did not name this category explicitly, and the absence of explicit naming was being exploited.

So: a ban, with teeth, on a reading of the November 2025 negotiating drafts that earlier seemed unlikely to survive. I will take it. I want to be precise, though, about why I think it made it through.

Bans are easier when the lobby on one side is loud and the lobby on the other side does not really exist as a lobby.

The developers of these apps are not represented in Brussels. They have no trade association, no policy team, no MEP they have cultivated over three legislative cycles. The harms are documented in ways that are difficult to spin. The political cost of being seen to protect this category was higher than the political cost of banning it. When the asymmetry runs that way, even a deregulatory package can absorb a ban without much friction. The nudification clause did not survive because European institutions discovered new resolve on gendered harm. It survived because nobody who mattered fought for it to die.

The trade, named plainly. Industry got eighteen months on the high-risk regime. The public got a ban on a category of app that should have been banned in the original Act and was not, because in 2023 and 2024 the harm had not yet acquired enough political salience. The legislative cycle has, in effect, used the salience of one harm to absorb the political cost of delay on the other. This is not a coincidence of timing. This is how omnibus packages work. You bundle something defensible with something that needs defending, and the package gets through.

I do not think the trade is straightforwardly bad. The nudification ban is real and will protect real people. The high-risk delay is real and will leave real people subject to unaccountable systems for an extra year and a half. Both are true. The thing I object to is the framing that treats them as equivalent gains, or as a balanced bargain between competing interests. They are not equivalent and the interests were not competing. The firms got a substantive concession on the regime that affects their costs. The girls and women got a clause that closes a specific loophole on a specific harm. Both deserved the protection. Only one of them had to wait for the other's concession to get it.

What "delay" means on the ground. I want to be specific about who is inside the eighteen months, because in the abstract it sounds like procedural housekeeping.

A worker in Poland whose CV is screened by an AI recruitment tool that the vendor claims has been debiased. Under the original timeline, that vendor would, by August 2026, have had to maintain documentation of how the debiasing was tested, on what populations, with what residual disparate impact, and would have had to make that documentation available to the national authority on request. Under the new timeline, that obligation begins in December 2027. The worker's CV is still being screened. The vendor still says the tool is debiased. The Polish authority still cannot demand the working.

A parent in Belgium whose child's school uses an AI system to predict which students are at risk of dropping out. The system flags the child. Under the original timeline, the provider would, by August 2026, have had to provide deployers (the school, the education authority) with clear information about the system's intended purpose, its accuracy on relevant populations, and known limitations. Under the new timeline, the school is free to use the flag, and the parent is free to wonder what it is based on, for another eighteen months.

A migrant at a Greek border crossing whose biometric data is processed by an AI identity verification system. The original timeline would have brought that system under the high-risk regime in August 2026. The new timeline says December 2027 if standalone, August 2028 if integrated into border management products. The migrant is, in both cases, subject to the system today. The difference is whether the system's provider has to demonstrate, by a specific date, that it works on people who look like the migrant.

These are not hypotheticals. The systems exist. They are deployed now. The Act's high-risk obligations were the thing that was supposed to convert "deployed" into "deployed accountably". The conversion has been postponed.

What I am watching. Three things, over the next year.

First, whether the December 2026 nudification deadline is enforced or whether enforcement gets quietly deprioritised once the political moment passes. The clause is only worth what the enforcement is worth. App stores have to delist. Hosting providers have to take down. National authorities have to act on reports. None of that is automatic.

Second, whether the harmonised standards actually arrive during the extended runway. If December 2027 comes and the standards are still incomplete, there will be another Omnibus, and another delay, on the same grounds. The delay only makes sense if the time is used. The political economy of standard-setting suggests the time will not be used as quickly as the deadline assumes.

Third, whether the precedent of "we bundled a protection with a delay" becomes the template for the next round. The AI Act has a review clause. There will be more Omnibuses. If the pattern holds, every future deregulation will arrive wrapped in a defensible specific harm, and the trade will get harder to refuse each time. That is how regulatory regimes get hollowed out, not through repeal, but through bundling.

I am glad the nudification ban is in the package. I am not glad about the price. And I want to be honest that calling it a price at all is already a concession to the framing I am trying to resist, because the protection of girls and women from image-based sexual abuse should not have been a bargaining chip in the first place. It was. That is the part worth remembering when the next package arrives.

Footnotes

¶