Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Regulation

European Parliament votes to delay AI Act high-risk rules

On March 26, the European Parliament adopted its position on the “digital omnibus” proposal that amends the EU Artificial Intelligence Act, by 569 votes to 45 with 23 abstentions. Parliament’s announcement says the changes delay certain rules on high-risk AI systems so that guidance and standards are ready to help companies comply, and add a new ban on so-called nudifier apps.

This is Parliament’s negotiating position, not the final law. Talks with the Council of the EU, which adopted its own position on March 13 according to Covington’s Inside Privacy, can now begin.

/ 01What Parliament voted for

The omnibus is part of a simplification package the European Commission proposed on November 19, 2025. Parliament’s version sets fixed dates rather than leaving timing open:

  • December 2, 2027 for high-risk systems listed in the Act, such as those used in biometrics, critical infrastructure, education, employment, law enforcement and border management.
  • August 2, 2028 for AI systems covered by EU sectoral safety and market surveillance legislation — products such as medical devices, radio equipment and toys.
  • November 2, 2026 for providers to comply with rules on marking AI-generated audio, image, video or text with its origin.

The text adds a prohibition on systems that create or manipulate sexually explicit or intimate images resembling an identifiable real person without consent, with an exception for systems that have effective safety measures preventing such images. It allows providers to process personal data to detect and correct bias, but only where strictly necessary. It extends support measures for small and medium enterprises to small mid-caps, and allows lighter AI Act obligations for products already regulated under sectoral laws.

Biometric Update notes that the delay follows the Commission missing its February deadline for guidelines on high-risk AI. Inside Privacy reports that both Parliament and Council would remove the Commission’s power to bring the dates forward, and that Parliament’s version reinstates registration in the EU database for systems that providers self-assess as not high-risk, which the Commission had proposed dropping.

/ 02Why it matters

The AI Act’s general application date is August 2, 2026. Without the omnibus, high-risk obligations would start applying on the original schedule even though the harmonised standards and guidance meant to explain how to meet them are not finished. Fixed later dates, if they survive negotiations, give providers and deployers a clear target and remove the uncertainty of a moving deadline.

It would be a mistake to read the vote as a retreat from the obligations themselves. The requirements for high-risk systems — risk management, data governance, technical documentation, record-keeping, transparency to deployers, human oversight, accuracy and robustness — are unchanged in substance. What changes is the time available to build them. Organisations that use the extra time to postpone the work will be in the same position in late 2027 that many were in this spring.

There is also a timing risk. Inside Privacy reports that trilogues were expected to start in April with an agreed text targeted for May. With the general application date of August 2 approaching, there is little slack: the changes have to be agreed and formally adopted before the dates they are meant to replace arrive.

/ 03Documentation is the hard part

The obligations that take longest to meet are the ones about evidence: being able to show what data a system was built on, how it reaches its outputs, and what happened in a given case. Technical documentation and automatic logging are not features that can be added in the final quarter before a deadline. They depend on how the system is designed.

Systems that reason over an explicit knowledge model have an advantage here. When inputs are mapped to a defined ontology and each output is derived from identifiable facts by rules that can be inspected, the documentation largely describes the system as it is, and the logs record which facts led to which result. When outputs come from a model whose internal reasoning cannot be inspected, the provider has to construct that evidence around it. Either way, the regulation asks for the same thing: a decision trail someone else can follow.

A delay changes when the AI Act asks for evidence, not whether it will.

/ 04What to watch or do

  • Trilogue outcome. The final dates and the scope of the nudifier ban will be set in negotiations with the Council. Plan against the Parliament and Council positions, which already agree on the main high-risk dates.
  • Formal adoption before August 2. Track whether the amendments are agreed and adopted before the Act’s general application date.
  • Use the time. Inventory AI systems, classify them, and start on documentation and logging now. Standards will refine the details, but the structure of the obligations is known.
How it unfolded
MAR 13, 2026Council adoptsits position[3]MAR 26, 2026Parliamentvotes 569–45[1]MAR 28, 2026This postAUG 2, 2026AI Act generalapplication[1]UPCOMINGNOV 2, 2026AI contentmarkingdeadline[1]UPCOMINGDEC 2, 2027Listedhigh-riskrules apply[1]UPCOMINGAUG 2, 2028Sectoralhigh-riskrules apply[1]UPCOMINGMAR 13, 2026Council adopts its position [3]MAR 26, 2026Parliament votes 569–45 [1]MAR 28, 2026This postAUG 2, 2026 · UPCOMINGAI Act general application [1]NOV 2, 2026 · UPCOMINGAI content marking deadline [1]DEC 2, 2027 · UPCOMINGListed high-risk rules apply [1]AUG 2, 2028 · UPCOMINGSectoral high-risk rules apply [1]
Aktiver’s take — Government

The new dates buy time to build evidence, not a reprieve from producing it. Technical documentation and logs assembled after the fact describe a system nobody can actually inspect.

Map the Act’s obligations onto the system itself. Each high-risk use links to its Annex III category, each risk control to the article that requires it, each training dataset to its governance record, and each logged decision to the input facts and rule that produced it. A conformity assessor’s question, such as which outputs last quarter depended on a given dataset version, becomes a graph query whose answer does not change between runs. Deterministic reasoning helps here: a decision replayed for an auditor returns the same result with the same evidence. Ingestion Hub can turn the regulation text, internal policies and system documentation into one ontology, exported as OWL or RDF, so legal and engineering teams work from the same model.

How we’d model it

High-risk AI systemAnnex III categoryRisk controlAI Act articleLogged decisionInput factObligationApplication dateclassifiedUnderrequiredByderivedFromappliesFromHigh-risk AIsystemAnnex IIIcategoryRisk controlAI ActarticleLoggeddecisionInput factObligationApplicationdateclassifiedUnderrequiredByderivedFromappliesFrom
  • High-risk AI system —classifiedUnder→ Annex III category
  • Risk control —requiredBy→ AI Act article
  • Logged decision —derivedFrom→ Input fact
  • Obligation —appliesFrom→ Application date

See where this applies →

Sources — read the originals
  1. Artificial Intelligence Act: delayed application, ban on nudifier apps — European Parliament europarl.europa.eu ↗
  2. EU Parliament backs delaying AI Act deadlines for biometrics, high-risk applications — Biometric Update biometricupdate.com ↗
  3. MEPs Adopt Joint Position on Proposed Digital Omnibus on AI — Inside Privacy (Covington) insideprivacy.com ↗
Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading