EU AI Act: what changes on December 2, 2026 (checklist)

The dates, from December 2026 outward¶

The Digital Omnibus on AI rewired Article 113, the article that sets when each part of the Act applies. The result is a sequence of fixed calendar dates. The Commission AI framework page confirms the 2 December 2027 date for Annex III systems and 2 August 2028 for product-embedded systems, and describes the December 2026 prohibition on generated non-consensual intimate content.

Everything that applied on 2 August 2026 stays applicable. The August 2026 guide covers that layer in detail: Article 50 transparency for providers and deployers, Article 101 enforcement against general-purpose AI model providers, and the Chapter V model obligations in force since August 2025.

Change one: two new prohibitions with a penalty ceiling of 7%¶

The Omnibus added two practices to Article 5, both applicable from 2 December 2026. AI systems that generate or manipulate child sexual abuse material, and AI systems that generate or manipulate realistic sexual or intimate imagery of an identifiable person without their consent, are prohibited. The prohibition reaches a provider where that output is the intended purpose of the system or a reasonably foreseeable and reproducible result without significant technical modification.

Article 5 breaches sit in the top penalty tier of Article 99: up to EUR 35 million or 7% of total worldwide annual turnover, whichever is higher for an undertaking. That tier has applied to the original prohibitions since 2 February 2025; the two new categories join it on 2 December 2026.

The compliance work is concrete. A team shipping image, video, or audio generation needs adversarial test coverage against both categories, refusal behaviour that survives paraphrase and system-prompt pressure, and retained test results. A deployer embedding a third-party generator needs the same assurance in writing from the vendor.

Change two: the marking grace period ends¶

Article 50(2) requires providers of systems that generate synthetic audio, image, video, or text to mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated, as far as technically feasible. The duty applied on 2 August 2026 for systems placed on the market from that date. Systems already on the market before 2 August 2026 received one transitional period, which ends on 2 December 2026.

From 2 December 2026 the distinction between old and new systems disappears for this duty. Non-compliance sits in the Article 99(4)(g) tier: up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. The Article 50 checklist covers the full transparency control set, including the disclosure and deepfake-labelling duties that carried no grace period at all.

Provider checklist: this quarter (before 2 December 2026)¶

Work items for the provider role, ordered so that the items with hard December dates come first.

• Close out machine-readable marking on every generative system placed on the market before 2 August 2026. Verify the marking survives your real output paths: API responses, file exports, embedded viewers.
• Red-team the two new prohibitions. Run structured adversarial tests for CSAM and non-consensual intimate imagery generation on every model and system you place on the market. Keep the test plans and results.
• Re-check Article 50(1) disclosure on interactive systems: informed no later than first interaction, clear and distinguishable, in each language the surface serves.
• Record the evidence. For each control, keep a record that it ran: marking applied per output pipeline, disclosure served per surface, refusal behaviour per test cycle.
• Confirm your classification position. Record the Article 6 reasoning for each system, including any position that a system is out of Annex III scope. The risk classifier walks the Annex III logic.

Deployer checklist: this quarter (before 2 December 2026)¶

A deployer runs an AI system under its own authority. Most enterprises hold this role for most of their AI estate.

• Inventory generative surfaces. List every place your organisation publishes or distributes AI-generated media, including marketing assets produced with third-party tools.
• Get vendor commitments in writing for the 2 December 2026 marking close-out on any generative product that shipped before 2 August 2026, and for refusal behaviour on the two new prohibited categories.
• Keep the deepfake and public-interest text disclosures running. These deployer duties under Article 50(4) have applied since 2 August 2026 and continue unchanged; tie each disclosure to a release-time regression check.
• Check internal tooling. An internal image tool that can produce intimate imagery of an identifiable person is in scope of the new prohibitions even when no output is published.
• Log that controls fired. A record written at the time of the interaction answers a supervisor faster than a policy document.

Provider checklist: 2027 (to 2 December 2027)¶

Chapter III arrives for stand-alone Annex III systems on 2 December 2027, with the same substance it always carried: risk management (Article 9), data governance (Article 10), technical documentation (Article 11), logging (Article 12), transparency to deployers (Article 13), human oversight design (Article 14), accuracy and robustness (Article 15), and the Article 17 quality management system. The Omnibus deadline guide lists every date that moved.

• H1 2027: operate risk management and data governance with evidence, and make Article 12 logging capture the fields the Annex IV technical file will need.
• H1 2027: design human oversight into the product. Article 14 asks for a person who can interpret output, override it, and stop the system; that takes an interface and a record.
• H2 2027: assemble the Annex IV file and run the Article 17 QMS with internal audit evidence. Small mid-caps can use the simplified documentation forms the Omnibus extended to them.
• H2 2027: choose the conformity assessment route and, where a notified body is required, book it. Capacity does not grow with the deadline.
• By 2 December 2027: register under Article 49, including systems self-assessed as not high-risk under Article 6(3) with the reduced Annex VIII payload, and sign the declaration of conformity.

Deployer checklist: 2027 (to 2 December 2027)¶

The deployer duties in Article 26 and the Article 27 fundamental rights impact assessment follow the Annex III date.

• Classify your deployments against Annex III and record the reasoning. Creditworthiness assessment, insurance risk pricing for life and health, employment decisions, and essential-service access are the common enterprise triggers.
• Draft FRIAs early for in-scope systems. The assessment describes the deployment process, affected people, specific risks, oversight measures, and remedies; most of that content exists only once the oversight model is decided.
• Assign oversight owners per high-risk system, with authority to suspend use, and log their interventions.
• Verify vendor readiness. Ask each provider for its conformity assessment plan, instructions for use, and logging interface; Article 26 duties assume those inputs exist.
• Use existing DPIAs. Article 27(4) lets a GDPR data protection impact assessment be complemented rather than duplicated.

2028: product-embedded AI and the sectoral-overlap question¶

High-risk AI that is a safety component of a product covered by the Annex I harmonisation legislation, such as machinery, medical devices, or lifts, follows the sectoral conformity assessment already required for that product. Chapter III applies to those systems from 2 August 2028.

The Omnibus also addressed the overlap between the AI Act and sectoral product law. Analyses by Orrick, Gibson Dunn, and Cooley report that the Commission is empowered to limit AI Act requirements where sectoral legislation imposes equivalent obligations, with a delegated act on those overlaps due by 2 August 2028, and that AI systems covered by the Machinery Regulation are largely taken out of parallel AI Act scope. Those firms also report a deferral to 2 August 2030 for certain high-risk systems used by public authorities. Confirm the exact scope of each carve-out against the Official Journal text before relying on it.

• Providers of embedded AI: align the Chapter III work with the sectoral conformity cycle for the host product, and reuse the Annex IV structure across both files.
• Deployers buying regulated products with AI inside: ask which regime the vendor treats as leading, and get the 2028 plan in the contract.
• Both roles: track the sectoral-overlap delegated act; it determines how much duplicated documentation the 2028 wave actually requires.

Supervision and the small mid-cap category¶

Two structural Omnibus changes shape who you will deal with and under what regime. The Commission describes reinforced AI Office powers and centralised oversight of systems built on general-purpose AI models. Law-firm analyses add detail: the AI Office holds exclusive supervisory competence over AI systems built on a general-purpose model where the same provider or group supplies both, and over AI systems integrated into very large online platforms and search engines designated under the Digital Services Act, with powers that include investigations, inspections, binding commitments, and fines. For a GPAI-based system, the supervisory conversation moves to Brussels.

The Omnibus also extended the simplifications previously reserved for SMEs to small mid-caps. Orrick reports the thresholds as fewer than 750 employees and up to EUR 150 million turnover or EUR 129 million balance sheet total, drawing on Commission Recommendation (EU) 2025/1099, with simplified documentation templates, proportionate quality management, priority sandbox access, and proportionate penalties. An organisation near those thresholds should establish its status now, because it changes the Article 17 and Annex IV workload for 2027.

Evidence is the common thread¶

Each checklist item above ends the same way: show that the control ran on the day it mattered. The marking pipeline, the disclosure, the refusal, the oversight decision, and the classification reasoning all need a record produced while the work happened.

That is the layer KLA operates. The KLA Policy Engine evaluates a Decision Request before a governed agent action runs, Decision Desk holds the require_approval outcomes for a named human, and Evidence Room produces a Sealed Evidence Bundle a third party can verify offline. Building that layer during the December 2026 transparency work means the 2027 high-risk evidence exists as a by-product of operating, and the EU AI Act hub maps each obligation to the control that produces its record.