In brief: The AI labelling obligation under the EU AI Act, applicable since 2 August 2026, covers only deepfakes, direct AI communication with customers, and unedited AI-generated texts of public interest — but not most typical enterprise applications.
Since 2 August 2026, the EU AI Act has required, for the first time, binding transparency obligations for certain AI-generated content such as deepfakes and chatbot communication. For compliance officers, the question arises as to which use cases are actually affected — since there is no blanket labelling obligation for all AI content.
With effect from 2 August 2026, binding transparency obligations for AI-generated content apply for the first time under the European AI Act. The purpose of the rule is to protect consumers and business partners from misleading representations and to improve the traceability of AI use. The Regulation does not provide for a general labelling obligation for all content generated with AI, but instead ties the obligation to clearly defined use cases.
Accordingly, AI-generated or manipulated images, videos and audio recordings that depict real people, objects, places or events in a deceptively realistic manner — so-called deepfakes — are subject to the labelling obligation. Companies must likewise disclose when customers communicate directly with an artificial intelligence, for example via chatbots or virtual assistants on websites and in customer portals. A narrower condition applies to AI-generated texts: labelling is only required if the text is published without human review and serves to inform the public on matters of public interest.
For many typical enterprise applications, the labelling obligation therefore does not apply from the outset. Product descriptions, marketing texts, translations, internal documents or the AI-assisted analysis of business data are regularly not affected. The same applies to mere linguistic polishing or automated spell-checking of already existing texts. In the case of automated text generation, the decisive factor is always whether a human reviews the content editorially before publication or assumes editorial responsibility and liability for the text — in this case, the special transparency obligations do not apply, and labelling can be omitted.
Likewise, not every AI-assisted editing of audiovisual content automatically triggers a labelling obligation. Anyone who merely extends image backgrounds, optimises colours, improves image quality or removes disruptive elements generally falls outside the scope of the provisions, since such creative adjustments are in principle not considered subject to labelling.
In practice, this means for compliance officers: instead of a blanket labelling of all AI content, a case-by-case assessment is required based on the three categories of deepfakes, direct AI communication, and unedited publication of text of public interest. Companies should document internal processes and responsibilities accordingly — in particular the editorial review and approval of AI-generated content — in order to be able to demonstrate human oversight if needed.
Source: www.it-daily.net · Published 14 August 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.