Digital Omnibus AI Regulation: How to Read the 4-Column Comparison Table

When the European institutions publish legislative comparison documents during interinstitutional negotiations, they follow a standardised format that can be opaque to readers unfamiliar with EU legislative procedure. The 4-column comparison table for the Digital Omnibus AI amendments is one of the most practically useful documents available to compliance professionals, but only if you understand what each column represents and how to extract actionable intelligence from tracked changes.

This guide explains the document's structure and walks through the most consequential tracked changes in the articles that matter most for enterprise compliance planning.

What the 4-Column Comparison Document Is

The 4-column comparison document is produced by the Council General Secretariat or the European Parliament's legal service to facilitate trilogue negotiations. It does not represent any institution's final position. It is a working tool that shows, in parallel, what each party brought to the table so that negotiators can identify where text is agreed, where it is contested, and what the realistic landing zones might be.

The four columns are structured as follows.

Column 1: Current EU AI Act text — this is the baseline, Regulation (EU) 2024/1689 as it stands after publication in the Official Journal. Any provision not proposed for amendment appears identically in all four columns, signalling consensus that it should remain unchanged.

Column 2: Commission's Digital Omnibus amendment — this reflects COM(2025) as published, showing which provisions the Commission proposes to modify and how. Text struck through in this column is proposed for deletion; bold or underlined text is proposed addition.

Column 3: Council's position — this reflects the general approach agreed by member states. Where the Council diverges from the Commission proposal, the text will differ from Column 2. Council amendments to Commission amendments — so-called "amendments to amendments" — are among the hardest passages to read in the document.

Column 4: Parliament's position — this reflects the mandate adopted by the relevant Parliamentary committees and confirmed in plenary. Parliament frequently proposes more substantive changes than either the Commission or Council, particularly on fundamental rights provisions.

When all four columns show identical text, that provision is effectively agreed and can be treated as settled law for compliance planning purposes.

Walking Through the Most Significant Tracked Changes

Article 3 — AI System Definition

Article 3(1) of the current EU AI Act defines an AI system as a machine-based system designed to operate with varying levels of autonomy that may exhibit adaptiveness after deployment, and that infers how to generate outputs such as predictions, content, recommendations, or decisions influencing real or virtual environments.

The Digital Omnibus proposal in Column 2 introduces a clarification that the inference requirement distinguishes AI systems from traditional software. The Council position in Column 3 adds an explicit recital cross-reference intended to exclude rule-based expert systems without machine learning components. Parliament's Column 4 text is closer to the Commission's version, with added language ensuring that systems designed to circumvent the definition through artificial architectural choices remain in scope.

For compliance planning, the definition debate matters because it determines whether legacy automation systems in HR or credit scoring are subject to AI Act obligations at all. Monitor Column 3's expert system carve-out closely.

Article 6 — High-Risk Classification

This is the most commercially significant contested provision. The current text in Column 1 establishes a two-part test: systems listed in Annex III are high-risk unless the provider demonstrates they do not pose significant risk of harm to health, safety, or fundamental rights.

The Commission's Column 2 text introduces a new stand-alone risk assessment methodology that gives providers clearer criteria for the significant risk determination. The Council's Column 3 text adds a further safe harbour for AI systems used purely for internal administrative support functions, provided they have no direct effect on individuals subject to the AI's outputs.

Parliament's Column 4 text rejects both modifications, maintaining the current Article 6 structure while adding clarifying recitals. The Parliament's position is that the Commission and Council modifications create loopholes that would allow high-risk applications to escape regulatory scrutiny through definitional engineering.

The practical implication is that until this article is resolved, providers and deployers of systems in Annex III categories cannot be certain whether a simplified or full conformity assessment applies.

Article 28 — Deployer Obligations

Article 28 sets out what organisations deploying high-risk AI systems must do, as distinct from providers who develop them. The current text requires deployers to use systems in accordance with instructions, ensure human oversight, monitor performance, and report serious incidents.

The Commission's Column 2 amendments clarify the boundary between provider and deployer obligations for systems used in cloud-based configurations. The Council's Column 3 text further reduces deployer documentation obligations where providers have already supplied full technical documentation — a practical accommodation for enterprise procurement. Parliament's Column 4 text strengthens deployer obligations by extending them to systems deployed in employment contexts, with specific provisions on worker notification and rights to contest AI-driven decisions.

Article 43 — Conformity Assessment

Article 43 specifies whether a high-risk AI system requires third-party certification or can be self-assessed. The current text allows self-assessment for most Annex III categories except biometric identification systems and systems used in critical infrastructure.

The Commission's Column 2 text expands third-party assessment requirements to AI systems used in recruitment and education, citing documented evidence of discrimination in self-assessed deployments. The Council's Column 3 text maintains the current self-assessment model for the majority of categories, arguing that third-party assessment infrastructure is not yet sufficiently developed across member states to mandate it. Parliament's Column 4 text goes further than the Commission, requiring third-party assessment for all Annex III point 4 (employment) and point 5 (essential services) systems.

Article 51 — GPAI Obligations

Article 51 triggers enhanced obligations for GPAI model providers whose models exceed the systemic risk threshold. The compliance implications of this article extend beyond model developers to every enterprise that deploys GPAI models in their products, since deployers need to understand what documentation and testing the model provider is required to supply.

The Commission's Column 2 text maintains the 10^25 FLOP compute threshold. The Council's Column 3 text proposes raising this threshold and adding a market share criterion. Parliament's Column 4 text proposes maintaining the compute threshold while adding a qualitative trigger based on user base and downstream integration into critical sectors.

Compliance Planning Based on the Table

Use the 4-column table as a risk matrix. Provisions where Columns 1 through 4 are identical represent certainty — invest in compliance now. Provisions where the Commission and Parliament agree but Council diverges are likely to resolve closer to the Commission/Parliament position, since Council tends to make concessions on fundamental rights protections in return for operational flexibility elsewhere. Provisions where all three institutions diverge represent genuine optionality — design modular compliance frameworks.

Frequently Asked Questions

Where can I access the official 4-column comparison document? These documents are published in the Council's public register (register.consilium.europa.eu) and through the European Parliament's legislative observatory. Search for the interinstitutional file number for the Digital Omnibus AI amendments.

Is the 4-column comparison document legally binding? No. It is a working document used to facilitate negotiations. Only the final text published in the Official Journal of the European Union is legally binding.

How often is the comparison document updated during trilogue? New versions are typically issued after each trilogue round. During active negotiations, this may be monthly. The Council register timestamps document versions, allowing compliance teams to track changes between rounds.

Sources

• Regulation (EU) 2024/1689 (EU AI Act), Articles 3, 6, 28, 43, 51
• European Commission, Digital Omnibus AI Proposal, COM(2025) (interinstitutional reference upon publication)
• Council of the European Union, General Secretariat, Guide to the Ordinary Legislative Procedure
• European Parliament, Legislative Observatory, Interinstitutional File for AI Act Amendments
• Council General Secretariat, Interinstitutional Style Guide for Comparison Documents
• European Parliament Research Service, "How to read EU legislative documents," PE briefing series