2026-05-07 - Digital Omnibus on AI: the Provisional Agreement of 7 May 2026

---

In the early hours of today, Thursday, 7 May 2026, the negotiators of the European Parliament and the Council reached a provisional political agreement on the Digital Omnibus on AI. The news was officially released through the European Parliament’s press release (ref.

20260427IPR42011

). It will be the subject of the press conference of co-rapporteurs Arba Kokalari (EPP, Sweden) for the IMCO committee and Michael McNamara (Renew, Ireland) for the LIBE committee, scheduled for 11:00 CEST in Strasbourg. This is a development that, only nine days ago, looked remote: the trilogue of 28 April had ended after twelve hours of negotiation without agreement, the press conference scheduled for the following day had been canceled, and several institutional and professional sources pointed to 13 May as the next available window. The closure of the file on this date is therefore a politically significant event before being a juridical one.

In this contribution, we offer a first reading of the political text, as it emerges from the official Parliament communication and from the legislative procedure

2025/0359(COD)

. It is a preliminary reading, because the consolidated text will be available only in the hours following the press conference and may contain refinements that escape us now. But the substantive picture is clear enough to support an initial orientation.

The political and procedural frame

The 7 May agreement closes — politically — the negotiating phase of the Commission’s seventh Omnibus package, presented on 19 November 2025 as

COM(2025) 836

. The institutional path has been compressed: Council position on 13 March 2026 (

press release 189/26

), joint IMCO/LIBE report A10-0073/2026 on 18 March, plenary vote in Parliament on 26 March (569 in favour, 45 against, 23 abstentions), first political trilogue immediately afterwards, second trilogue on 28 April closed without agreement, and — in a decision that surprised some observers — the reopening of negotiations culminating in today’s deal. All under the Cypriot Presidency, in view of the end of its term on 30 June 2026.

The next step is formal adoption by both co-legislators, which must be completed before 2 August 2026, the date of general application of the AI Act for high-risk system obligations as currently in force. Until that adoption, the original regime of the AI Act remains juridically operative: the distinction between lex lata and de lege ferenda is, at this stage, anything but formal and deserves to be kept in mind in any operational assessment.

The new application dates

The political core of the agreement confirms the dates already shared by the two co-legislators in their negotiating mandates. For high-risk AI systems under Annex III — biometrics, critical infrastructure, education, employment, law enforcement, border management — the application of obligations is postponed to 2 December 2027. For AI systems used as safety components in products covered by EU sectoral safety and market surveillance law (Annex I), the application is postponed to 2 August 2028.

A third element concerns transparency for AI-generated or manipulated content: the technical marking obligation (watermarking) is postponed to 2 December 2026. This is a compromise among the three positions on the table: the Commission had proposed 2 February 2027, the Parliament had reduced the period to about three months, indicating 2 November 2026, and the agreed text adopts a four-month transitional period starting from 2 August 2026. The application date of the remaining provisions of Article 50 — those imposing transparency on the deployer and on the user regarding interaction with an AI system or the synthetic nature of content — is not expressly indicated in the official communication, and on this point we will await the consolidated text.

The Annex I turning point

The point on which the trilogue of 28 April had failed concerned the conformity assessment architecture for AI systems integrated as safety components in products already regulated by Union sectoral safety law (Machinery Regulation, Medical Devices Regulation, In Vitro Diagnostic Medical Devices Regulation, General Product Safety Directive, and the like). The political text expressly addresses machinery products; any extension to further categories will be clarified by the consolidated text. The Parliament’s position was clear in calling for a transition to the sectoral-only regime for products in Section A of Annex I; the Council resisted, concerned about fragmenting the AI Act’s horizontal framework into a plurality of parallel sectoral logics.

The 7 May agreement closes with a solution that aligns with the Parliament’s orientation, but with an equivalence clause. On this point, the political text exits the dual conformity regime for machinery products: for this category, the layering between AI Act requirements and sectoral safety law ceases, and the regulatory center of gravity shifts to the sectoral discipline, subject to the equivalence clause. The latter binds the sectoral discipline to ensure a level of health and safety protection no lower than that which the layered regime would have ensured.

In our reading, the scope of this passage will depend entirely on how the equivalence clause is articulated in the consolidated text and the implementing guidelines. An equivalence clause drafted in robust terms — with verifiable criteria and intervention powers — can preserve the level of protection; a clause formulated in merely programmatic terms risks being applied symbolically. This is a point we will return to once the text is available.

The narrowing of the “safety component” concept

The agreement also introduces a significant narrowing of the concept of safety. The perimeter of the notion is narrowed by introducing a functional threshold: AI functions that exhaust themselves in user support or performance optimization exit the high-risk regime, where their potential malfunction does not translate into a concrete risk to health or safety. What emerges, in substance, is a causal test relative to harm as the qualifying element of the safety component.

In our view, this clarification was needed. The previous, broad formulation ended up drawing into the high-risk regime ancillary AI functions whose contribution to risk was marginal or absent. However, the new test requires non-trivial interpretive work: distinguishing when an AI function is merely assistive from when it contributes in a non-eliminable way to the safe operation of the product will be a task for providers, notified bodies and — ultimately — case law. We expect significant interpretive divergences in the early years of application.

Bias detection, personal data, fundamental rights

The agreement confirms the possibility of processing personal data — including special categories — where strictly necessary for the detection and correction of bias, both in high-risk and non-high-risk systems. The wording, as reported in the official communication, retains the strict necessity standard and provides for proper safeguards. This is a position consistent with what the EDPB and the EDPS argued in their

Joint Opinion 1/2026

on the Digital Omnibus, where they had asked not to weaken the protective regime.

The extension to all providers and deployers, regardless of system risk classification, was a choice that had drawn criticism from civil society organizations. CDT Europe and over forty organizations had signed a letter to the Parliament in mid-April 2026 flagging the risk of weakening fundamental rights protections, particularly for biometric identification systems, AI use in schools, and medical AI. In the preliminary reading we offer here, the point of balance will once again depend on the concrete articulation of strict necessity and proper safeguards: without verifiable criteria, the extension shifts the center of gravity of the balancing onto the controller’s self-assessment.

The ban on nudifiers and AI-generated CSAM

Among the most significant innovations of the agreement is the ban on AI systems that generate child sexual abuse material (CSAM) or that depict the intimate parts of an identifiable person, or that represent that person engaged in sexually explicit activities, without their consent. This is a new prohibition inserted within Article 5 of the AI Act, taking up — and broadening — the position introduced by the Parliament during the negotiating mandate.

The architecture of the prohibition is tripartite and covers the full commercial and operational life cycle of the system. On the supply side, placing on the Union market is prohibited in two variants: where the system is designed with the specific purpose of producing the prohibited content, and where — even absent such purpose — it lacks safety measures reasonably suitable to prevent it. On the use side, the prohibition applies directly to the deployer whenever such systems are used to generate the prohibited content. The material scope extends to images, video and audio. The compliance deadline for businesses is set at 2 December 2026.

This is — whatever assessment one wishes to make of the overall direction of simplification of the Digital Omnibus — a provision that fills an objective gap: Article 5 in its original version did not expressly prohibit this specific type of system, and enforcement proceeded through a combined application of GDPR, national criminal legislation and DSA, which left margins of uncertainty. The tripartite structure of the prohibition — purposive placing, placing without safeguards, deployer use — is moreover interesting because it distinguishes the provider’s plane from that of the deployer, in continuity with the mapping of the AI Act’s subject roles which we analyzed in our study

arXiv:2510.13591

The other changes: SMC, GPAI, simplifications

The agreement then confirms a series of provisions that had already been consolidated during the negotiating process. The extension of SME exemptions to small mid-cap enterprises (SMC) to accompany their growth beyond the SME size threshold without suddenly exposing them to the full set of obligations. The streamlining of enforcement of certain GPAI systems within the AI Office, which strengthens the role of the new European regulator as a central authority for general-purpose artificial intelligence models.

Other elements remain in force — in continuity with the position shared by the two co-legislators — including the simplified registration obligation in the European database for systems self-classified by providers as non-high-risk, a provision that the Parliament and the Council had reinstated relative to the Commission’s original proposal in line with the EDPB and EDPS suggestion, and the extension of the AI Office’s powers for the supervision of AI systems built on GPAI models.

A reading: the last-minute rescue and what it reveals

The way in which this agreement was closed is itself a juridically relevant fact. In our view, two elements warrant isolation.

The first concerns the fact that the rupture of 28 April and its composition of 7 May did not occur by accident. The breaking point — conformity assessment for AI integrated in Annex I products — is the manifestation of a structural node that the debate on the qualification of the AI Act’s subject operators had already made visible. When an AI system is a safety component of a product already regulated by sectoral norms, the qualifications of the AI Act provider, the sectoral manufacturer, and the final deployer come into tension with the regulation’s horizontal architecture. This is a terrain in which the definitions of Article 3 and the role transformation mechanisms of Article 25 show their operational limits — a point we discussed in our paper on

Subject Roles in the EU AI Act

, where the mapping of the six AI Act operators shows how the overlap between horizontal roles (AI Act) and vertical roles (sectoral regulations) generates zones of indeterminacy that the legislator could not definitively resolve.

The second element concerns the gap between the political framing of “simplification” and the actual juridical outcome. The instrument was presented as a measure of simplification and competitiveness; in its final version, it introduces a new general prohibition under Article 5 (nudifiers and CSAM), a causal test for the safety component concept, an equivalence clause for machinery products, a new legal basis for processing of special categories of data for bias detection purposes, a modulation of the penalty regime for SMCs, a redefinition of the AI Office’s powers. The overall complexity of the system, in our reading, does not diminish: it is redistributed. “Simplification” is therefore a political label that requires, from the jurist’s standpoint, to be disaggregated into its specific effects.

What changes for deployers (and what does not)

On the operational level, it is essential to keep two planes distinct. The first is that of lex lata. Until the formal adoption of the Omnibus by both co-legislators and publication in the Official Journal, the AI Act remains in force in its original wording. As of today, 2 August 2026, the date on which obligations for Annex III high-risk systems become applicable remains 2024, as written in the 2024 text. Organizations planning compliance against the new date of 2 December 2027 are doing so de lege ferenda, accepting the risk that formal adoption may not be completed in time or that the final text may contain specific modifications relative to today’s political text.

The second plane is that of de lege ferenda: once the amending regulation is adopted and published, deployers of Annex III high-risk systems will have a further deadline up to December 2027, and deployers of systems integrated in Annex I products will have until August 2028. The transition from lex lata to de lege ferenda planning is not automatic: it must await the formal juridical fact.

In any case, what continues to apply — because the Digital Omnibus does not touch it — are the obligations already in force: the Article 5 prohibitions (with the addition of the new ban on nudifiers and AI-generated CSAM, applicable from 2 December 2026), the AI literacy obligation under Article 4, the obligations on GPAI models applicable from 2 August 2025, and the governance provisions. For those who wish to delve deeper into the deployer’s individual obligations, we refer to our

operational checklist on Article 26

and to the recent piece on

DPIA and FRIA

Conclusion

The 7 May agreement closes — on the political level — a compressed and high-tension negotiation. As co-rapporteur Kokalari summarised, politics shows it can move “

just as quickly as technology

”. It remains to be seen whether the speed of the process will be reflected in the quality of the consolidated text, and whether the compromise formula on Annex I — the equivalence clause — will hold under the test of concrete application.

Systematically, the Digital Omnibus on AI emerges from the trilogue as an act that, behind the simplification label, redesigns the distribution of obligations within the AI Act ecosystem in a non-marginal way. The complexity of the regime does not decrease: it is recomposed. This is a recomposition with evident immediate beneficiaries — the machinery and regulated products industry, scaling SMEs, GPAI providers who see enforcement streamlined — and less visible counterweights: the AI Act’s horizontal framework, which exits a little more jagged, and — according to many civil society organizations — the robustness of protections in sectors where the pre-existing sectoral discipline is less strong.

At the operational level, the message is clear: those planning compliance cannot yet wind down their adjustment program to the current regime. The 2 August 2026 wall remains legally in force until the formal adoption of the Omnibus, and professional prudence requires building compliance programs under the original regime, treating any future relief as a possible supervening benefit rather than a planning assumption.

We will continue to follow the file. The next relevant step — the consolidated text of the agreement, the positions of the Commission and Council in the coming hours, and the formal adoption acts — will be the subject of a dedicated update.

---

Main institutional sources