AI Act: The postponement of high-risk cases does not constitute a general reprieve

GAIGI Team

— 3 min de lecture
Partager

August 2, 2026, does not mark a general moratorium. The Omnibus Act postpones certain high-risk requirements, but the transparency provisions of Article 50, general-purpose AI models, and enforcement mechanisms remain critical.

The Essentials in 30 Seconds

  • Annex III: December 2, 2027.
  • Annex I: August 2, 2028.
  • Article 50: Transparency.
  • Labeling transition: December 2, 2026.

What Happened

Article 50 requires disclosure during direct interaction with AI and machine-readable labeling of synthetic content. Deployers must inform individuals exposed to emotion recognition or biometric categorization, and disclose hyper-manipulations under the specified conditions. The transition effective December 2, 2026, applies to synthetic content systems placed on the market before August 2.

The high-risk timeline must be read in conjunction with the provisions that take effect on August 2. Organizations cannot assume that no obligations apply to them prior to the new dates. They must first determine whether they provide, deploy, or integrate a system covered by the existing rules.

Informing individuals serves a different purpose than technical labeling. The former pertains specifically to direct interaction with an AI system or certain processing operations. The latter aims to enable the identification of synthetic content. Therefore, responsible parties must provide distinct responses tailored to the context of use.

The transition period for certain systems already on the market does not eliminate the need to plan ahead. It establishes a specific date for the labeling of these systems. New versions and deployment scenarios must be tracked separately to avoid applying an exception more broadly than intended.

Analysis

Providers of general-purpose AI models must maintain technical documentation, provide information to integrators, comply with copyright laws, and publish a summary of the training content. The Commission entrusts the AI Office with the oversight of Chapter V. National authorities may access the documentation and, under certain conditions, the training, validation, and testing data, as well as the source code. The postponement does not eliminate these controls.

The framework applicable to general-purpose AI models also structures the relationship between the model provider and the parties that reuse it. The information intended for integrators must enable them to assess the model’s capabilities and limitations within a specific system. A generic guide is not always sufficient to inform the chosen deployment.

Technical documentation serves several purposes. It helps the provider explain its product, the integrator design its service, and the regulatory authority understand the choices made. Its quality depends, in particular, on the consistency between the described model, the relevant data, and the stated control mechanisms.

The role of the AI Office confirms that postponing the chapter on high risk does not eliminate all oversight capabilities. Teams must therefore be prepared to produce organized information. Waiting until the very last minute may make it more difficult to reconstruct decisions.

Operational Impact

For each system, the decision must be linked to a specific purpose, an identifiable version, a legal role, the individuals likely to be affected, and the deployment context. An up-to-date inventory, controlled user instructions, and a clear assignment of responsibilities are prerequisites for demonstrable compliance. Management must be able to distinguish between directly applicable obligations, deferred deadlines, and tasks that remain necessary in all cases. This discipline facilitates responses to regulatory authorities, contract management, and coordination among business teams.

The first step is to separate systems that generate or distribute synthetic content from other tools in the portfolio. For the former, it is necessary to verify where the user receives the information, how the content is labeled, and which party controls the interface. These answers may vary between an internal service and a product intended for the public.

For general-purpose models, the supplier-integrator relationship must be documented. The integrator needs to know the information received, the restrictions specified, and the version actually in use. This documentation becomes essential when a model change alters the characteristics of the final service.

A documentation review exercise is useful. The organization can select a system, gather documentation, deployment decisions, and transparency elements within a set timeframe, and then identify any gaps. The result transforms an abstract obligation into a concrete improvement to the system.

What to Do Now

  • Map — identify the system, version, purpose, and legal role.
  • Document — preserve the reasoning and evidence.
  • Plan — assign a lead, set a deadline, and schedule a review.
  • Monitor — reconcile contracts, instructions, and technical measures.

Identify the pathways through which a person can interact with AI without clearly identifying it. Teams must specify which message is displayed, at what point, and in which language. The review must cover both primary interfaces and backup paths.

Develop a system for managing synthetic content. It must assign responsibility for labeling, provide for pre-release testing, and document any technical limitations identified. Decisions must be documented when content is distributed across multiple channels.

Request the necessary integration documentation from model providers, then file it alongside the model version. This process ensures that the information provided is actually used in internal assessments and in the instructions given to teams.

Sources

  • AI Act Service Desk, Article 50 of Regulation (EU) 2024/1689 on transparency obligations: ai-act-service-desk.ec.europa.eu
  • European Commission, AI regulatory framework: General application as of August 2, 2026, and new high-risk dates: digital-strategy.ec.europa.eu
  • European Parliament, Legislative Timeline: Postponement of labeling to December 2, 2026, for systems placed on the market before August 2, 2026: europarl.europa.eu

Lire la suite