Digital Omnibus and the EU AI Act's Prohibited AI Practices: What Changed and What Did Not

What the Digital Omnibus changed in Article 5

The Commission proposed the Digital Omnibus on AI on 19 November 2025 as COM(2025)0836, with amendments to Regulations (EU) 2024/1689 and (EU) 2018/1139. Parliament adopted its amendments on 26 March 2026. The Commission reports a political agreement on 7 May 2026. The final act is Regulation (EU) 2026/1744 of 8 July 2026, which amends the AI Act, Regulation (EU) 2018/1139 and Regulation (EU) 2023/1230. The proposal is therefore superseded, and the open question is no longer whether the Omnibus touches Article 5 but how.

It does. The regulation amends Article 5(1), first subparagraph, points (ba) and (bb). The Commission now states that the AI Act prohibits nine practices: prohibitions 1-8 took effect in February 2025, and prohibition 9 was introduced through the Omnibus and comes into effect in December 2026.

A team that assumed Article 5 was untouched by the Omnibus, because other parts of the Act saw timeline changes, needs to correct that assumption. A team that assumed the existing prohibitions were rewritten should also correct it, because nothing reviewed shows any change to prohibitions 1-8.

• Changed: Article 5 points (ba) and (bb) were added by Regulation (EU) 2026/1744, described by the Commission as prohibition 9.
• Unchanged as far as reviewed: Prohibitions 1-8, which have applied since 2 February 2025.
• Superseded: The November 2025 proposal and Parliament's March 2026 amendments, replaced by the adopted regulation.
• Date to track: Prohibition 9 takes effect in December 2026.

What prohibitions 1-8 still require

Prohibitions 1-8 are a standing obligation. They include manipulative techniques, social scoring, and emotion recognition in workplace and education settings, among others. They have applied since 2 February 2025, alongside the AI literacy obligations, ahead of the Act's general application date. The Commission has also published guidelines on prohibited AI practices, with legal explanations and practical examples for stakeholders.

For compliance leaders, the practical reading is simple. The Omnibus gives no basis to relax, pause or remove controls built for prohibitions 1-8. Those controls should be mapped to the official Article 5 text and to the Commission guidelines, not to secondary summaries. The guidelines' examples are useful as test cases, because they describe the kinds of behaviour a control has to catch.

The obligation applies to the systems as deployed. A policy document stating that prohibited uses are not permitted is not evidence that they are not occurring. Evidence has to come from what the deployed systems actually did.

The new prohibition: purpose-based scoping for providers and deployers

The amended Article 5 treats providers and deployers differently. For providers, placing on the market or putting into service an AI system that generates or manipulates the material or performance described in points (ba) or (bb) is prohibited only where specified conditions are met. The first condition is that the generation or manipulation is the intended purpose of the system. For deployers, use of such a system is prohibited only where the deployer uses it for the purpose of generating or manipulating that material or performance.

This makes the new prohibition different in structure from a prohibition defined by outcome alone. Purpose is part of the test. A compliance programme therefore needs to record, for each system, the declared intended purpose and, for each deployment, the deployer's purpose of use. It also needs a clear record of whether the organisation acts as provider, deployer or both for that system.

The final conditions should be confirmed against the full text of Regulation (EU) 2026/1744 before controls are finalised, with legal sign-off on the interpretation. Treat the new prohibition as its own policy domain with its own effective date, separate from prohibitions 1-8.

Audit-ready evidence and where Trussed fits

An auditor or regulator will ask what the system was permitted to do and what it did. The useful evidence is specific: for an interaction or tool call, what was requested, which policy and version applied, whether it was allowed, redacted, blocked or escalated, and who approved any escalation. Sampled reviews of both blocked and allowed events show that controls operate on live traffic and are not only configured. Records of role allocation, purpose tagging and policy change history complete the picture.

This evidence has limits. Runtime records cover what the enforcement layer observes. They do not by themselves establish legal classification, and interpretation of Article 5 stays with legal and compliance owners.

Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with outcomes of allow, redact, block, rate limit or escalate for human approval at the level of individual interactions and tool calls. Enforcement generates continuous audit-ready evidence mapped to the EU AI Act, among other frameworks. Trussed deploys self-managed in the customer cloud, with configurable retention and data residency. It has run for 18+ months in continuous production in a regulated enterprise.