Digital Omnibus on the AI Act: what changes – and what does not

Chapter III Sections 1 to 3 of the AI Act – classification, requirements for high-risk systems and the obligations of providers, deployers and other operators – apply to Annex III systems from 2 December 2027 and to Annex I systems from 2 August 2028. The deployer obligations in Art. 26 move with them, for example keeping logs for at least six months and informing workers’ representatives in advance.

The rule for legacy systems matters in practice (Art. 111(2)): what is placed on the market or put into service before the date of application is covered by the high-risk obligations only if its design changes significantly afterwards. Systems intended for public authorities must comply by 2 August 2030 regardless. Recommendation: record the date each system went live and any later changes – that is the only way to show later whether a system benefits from the legacy rule.

Until now, providers and deployers had to take measures to ensure, "to their best extent", a sufficient level of AI literacy among their staff. The new text requires measures to support the development of AI literacy – taking into account knowledge, experience, training and the context of use. Paragraphs 2 and 3 are new: the Commission and Member States promote implementation, the Commission publishes practical examples and the AI Board issues recommendations.

So the duty has not gone away; it is worded more softly. The AI Act itself provides no separate fine bracket for Art. 4 (Art. 99(4)). Since the Digital Omnibus (Regulation (EU) 2026/1744), however, Member States can penalise any infringement (Art. 99(1) as amended); in Germany Art. 4 is currently not subject to fines (§ 15 KI-MIG). Recommendation: keep documenting training, guidance and approval rules – it is the simplest evidence that you took measures, and high-risk deployers need it anyway because human oversight must be assigned to competent persons (Art. 26(2)).

From 2 December 2026, AI systems are prohibited that generate or manipulate realistic image, video or audio content depicting the intimate parts or sexually explicit acts of an identifiable person without that person’s free, specific consent (point (ba)). Systems that generate or manipulate child sexual abuse material within the meaning of Directive 2011/93/EU are prohibited too (point (bb)).

Paragraph 1a draws the line: for providers, the ban applies where such content is the system’s purpose, or a foreseeable and reproducible result of its design and functions without significant technical changes – and the system lacks reasonable safeguards. For deployers it applies where they use the system for such content. Recommendation for providers of generative AI: document filters and abuse protection by December.

The transparency duties apply as originally planned since 2 August 2026: notice of direct interaction with AI (para. 1), machine-readable marking of synthetic content (para. 2), information on emotion recognition and biometric categorisation (para. 3), disclosure of deepfakes (para. 4).

The Omnibus only adds a transition: generative systems placed on the market before 2 August 2026 have until 2 December 2026 for marking under para. 2. Para. 7 on codes of practice was also reworded. The code of practice on marking AI-generated content has been final since 10 June 2026; it is voluntary.

"SME" (Recommendation 2003/361/EC) and "small mid-cap" (Recommendation (EU) 2025/1099) are now defined. Small mid-caps are companies that have outgrown the SME thresholds but are not yet large; the exact thresholds are in the recommendation.

Both groups may provide technical documentation through the Commission’s simplified form (Art. 11), implement the QMS proportionately (Art. 17(2)) and benefit from the lower amount for fines (Art. 99(6) and (6a)). SMEs without partner or linked enterprises may meet certain QMS elements in a simplified way (Art. 63). The level of protection does not drop – the form is simplified, not the requirement.

The new Art. 4a exceptionally allows providers of high-risk AI to process special categories of personal data to detect and correct bias. Paragraph 2 extends this, under the same conditions, to providers and deployers of other AI systems and models and to deployers of high-risk AI where the bias could affect health, safety, fundamental rights or lead to prohibited discrimination – it does not create a duty to test for bias.

The conditions are strict: not achievable with other data, including synthetic or anonymised data; pseudonymisation and limits on re-use; strict, documented access control; no transfer to third parties; deletion after correction or at the end of the retention period; reasons recorded in the record of processing activities. Art. 2(7) makes clear that the GDPR otherwise remains unaffected.

Deployers who must carry out a fundamental rights impact assessment – public bodies, private entities providing public services and deployers of systems under Annex III point 5(b) and (c) – may refer to the relevant sections of their DPIA for duties already met by it.

Previously the text only said the FRIA "complements" the DPIA. The AI Office template questionnaire, also as an automated tool, is now expressly to offer the option of including such references. Recommendation: run DPIA and FRIA in one joint process so that references are easy to trace.

For AI systems based on a general-purpose AI model where model and system come from the same provider or group, the AI Office is now exclusively competent. The same applies to AI systems that constitute or are integrated into a very large online platform or search engine under the Digital Services Act. Exceptions include product AI under Annex I and systems under Annex III point 2.

The AI Office gets its own powers: requests for information, inspections, binding commitments, fines under Art. 99 and periodic penalty payments of up to 5% of average daily turnover per day. For deployers, this competence applies only if they are also the provider or part of the provider’s group.