Has the entire AI Act been postponed?

No. The timetable is phased. Transparency applies from 2 August 2026; the high-risk dates differ by category.

Does a website score certify legal compliance?

No. Automated public checks are indicative and do not replace a legal review of actual uses, contracts and responsibilities.

AI ActAugust 1, 2026 9 min

AI Act: transparency from 2 August 2026 and the high-risk timetable

The AI Act applies in stages. Transparency applies from 2 August 2026; distinguish Annex III on 2 December 2027 from Annex I products on 2 August 2028. Practical checks for SMEs, with official sources and human review.

Facebook X LinkedIn WhatsApp

AI-generated illustration of Bernard Frisque with robots, European AI rules — AI-generated illustration.

A phased timetable, not one universal deadline

The AI Act entered into force on 1 August 2024. Its application is phased: initial prohibitions and AI literacy from February 2025, general-purpose AI rules from August 2025, transparency from 2 August 2026. The Commission’s updated timetable distinguishes Annex III high-risk systems on 2 December 2027 and high-risk systems embedded in Annex I products on 2 August 2028. Do not describe 2027 as full general application.

Who is concerned?

A provider develops or markets an AI system; a deployer uses it professionally. Using a purchased AI tool does not make an SME its developer, nor does it remove all responsibilities. Classify the actual function, role, data and context before deciding which obligations apply. Purely personal use is distinct from professional deployment.

AI literacy and human supervision

People using AI need appropriate understanding of the tools, limitations, hallucinations, privacy risks and when human intervention is essential. Keep practical instructions and evidence of the measures taken. Do not sell an invented compulsory European certificate or imply that a standard training course guarantees compliance.

Disclose chatbots clearly

When a person interacts with an AI system, transparency information must be clear unless that interaction is already obvious in context. Check the first interaction, not just a buried privacy-policy paragraph. Provide an accessible route to a human contact for important questions.

Deepfakes and generated content

Distinguish a provider’s technical marking duties from a deployer’s public disclosure obligations. Deepfakes that can appear authentic require appropriate disclosure. Not every clearly fictional AI illustration automatically has the same legal status. Artistic or satirical contexts require a specific assessment, not a blanket exemption.

Public-interest articles and editorial responsibility

Generated or manipulated text on matters of public interest requires an assessment of the transparency rules and the human-review exception. Genuine editorial control means checking facts and sources, correcting errors and having a person take responsibility; proofreading alone is not substantive verification. Keep records of sources and approval.

HR and other high-risk uses

Recruitment, candidate ranking, education, credit, biometrics and access to essential services can require a high-risk analysis. Annex III and product-linked Annex I categories have different dates. Identify the tool, discrimination risks, supplier, data handling and human decision process now; a later date is not permission to ignore other applicable laws.

Prohibited practices and sanctions

Some harmful manipulation, social scoring, non-targeted facial-image scraping and emotion-recognition uses are already prohibited, subject to the regulation’s exact definitions and exceptions. Sanctions are statutory ceilings, not automatic fines or a website score. Obtain qualified advice before applying a prohibition or penalty to a particular organisation.

An SME checklist

Inventory the AI tools, providers, purpose, users and data. Define prohibited inputs and human checks. Review public content and chatbot notices. Identify sensitive uses. Keep training, contracts and editorial evidence. Assign a human contact to coordinate priorities. An automated scan can support this work but cannot certify the organisation.

Computing capacity is a separate issue

European investment programmes for AI infrastructure concern computing capacity and development. They do not grant a business legal compliance, certification or partner status. Consult the Commission and EuroHPC sources separately from the AI Act application timetable.

Official sources and limits

General information, not legal advice. Verify the applicable text and your actual scope before acting.

Discuss your situation with Bernard →