Article 50: distinguish machine marks from public disclosure

The EU AI Act was adopted on 13 June 2024 and published on 12 July. Its original general application date was 2 August 2026, distinct from earlier application of Chapter I on 2 February 2025. Article 50 therefore concerned a future obligation when the original text was published, but that date is already past relative to this article’s publication on 25 October 2026. EUR-Lex also identifies a consolidated version dated 27 July 2026. Do not describe the original future timetable as an unchanged present-day compliance conclusion. 1

Article 50(1) in the original regulation concerns systems intended to interact directly with natural persons. Providers must design them so people are informed they are interacting with AI, unless this is obvious from the perspective specified in the text. This is not the same duty as marking a generated image. A chatbot introduction, a voice interaction and an avatar require different interface decisions. Test the information at the point of interaction, including accessibility and language, rather than relying on a notice buried in general terms.

Article 50(2) addresses providers of systems generating synthetic audio, images, video or text. The original text requires machine-readable marking and detectability, with effectiveness and technical feasibility qualifications. It also contains exceptions concerning standard editing assistance and specified lawful uses. A visible badge alone is not automatically a machine-readable solution. Conversely, metadata that a platform strips during upload may be invisible to ordinary readers. Procurement should ask what survives export, compression and redistribution, and what evidence supports the supplier’s claims under realistic processing conditions.

Article 50(4) in the original text assigns disclosure duties to deployers for deepfake image, audio and video content. Artistic, creative, satirical and fictional works receive a qualified disclosure approach, not a blanket claim that all art is exempt. The same paragraph addresses AI-generated or manipulated text published to inform the public on matters of public interest, with an exception involving human review or editorial control and editorial responsibility. That is more specific than saying every AI-assisted sentence must carry an identical label.

Create a content register identifying the tool, provider, publishing organization, media type and intended audience. Record who reviews public-interest text and who accepts editorial responsibility. Inspect transformation steps that can remove provenance data. These are operational recommendations rather than a legally exhaustive checklist. Company knowledge helps organize approved source material; provider obligations covers a different layer of the value chain. Do not confuse Article 50 output transparency with general-purpose-model documentation or with the literacy work discussed elsewhere in this batch.

The legal record identifies an amending regulation dated 8 July 2026, published on 24 July, and a consolidated text on 27 July. The original provisions above are a dated explanation, not a complete verification of every transition rule, later implementation measure or national enforcement position as of publication. A voluntary code or explanatory guideline is not identical to legislation. Use the EU AI Act overview for the broader framework, but obtain a version-specific assessment before asserting compliance. This article provides no legal advice and does not endorse any watermark as legally sufficient.