A photograph shows a chief executive at a press conference that never took place. The image was generated by artificial intelligence and looks entirely authentic. This is precisely the scenario the European Union had in mind when it drafted transparency rules that took effect on 2 August 2026 under the EU AI Regulation. Yet the obligations are narrower than many businesses assumed.
What must be labelled
The regulation imposes two main transparency duties. First, companies that publish AI-generated or AI-manipulated images, audio or video must label those contents whenever they convincingly depict real persons, objects, places or events, meaning a viewer could mistake them for authentic material. These are what the regulation calls deepfakes. The obligation applies regardless of whether the content is used in advertising, news or corporate communications.
Second, any business that lets customers interact directly with an artificial intelligence must disclose that fact. Chatbots on websites, virtual assistants in customer portals, and similar interfaces all fall within this requirement. The user must know they are exchanging messages with a machine, not a person.
Where the text exemption applies
The bigger question for most companies concerns text. The short answer is that routine AI-generated text is largely outside the scope of mandatory labeling. Product descriptions, marketing copy, translations and internal documents produced with AI assistance do not trigger a disclosure obligation, provided a human reviews or takes editorial responsibility for the content before publication.
Even without human review, text requires labeling only in a narrow circumstance: when it is published without oversight and serves to inform the public about matters of public interest. An e-commerce retailer auto-generating product listings, or a marketing team using AI to draft social media posts that a copywriter then edits, would not need to attach an AI label. A news organisation publishing AI-written articles on political developments without any editorial checks, by contrast, probably would.
The human review safeguard
The regulation draws a clear line around human involvement. If a person reviews the AI-generated content and takes editorial responsibility for it, the special transparency obligations under Article 50(4) of the AI Regulation fall away. No label is required.
This matters because it gives companies a straightforward compliance route. Rather than auditing every piece of content for its AI provenance, firms can simply ensure that a named individual reviews and approves material before it goes live. Frank Fischer, head of the IP and Media practice group in Germany at the law firm Rödl, has highlighted this distinction as central to understanding what the regulation actually demands of businesses.
Labeling does not equal legality
One point the regulation does not resolve is underlying legal compliance. A correctly labelled deepfake that infringes someone's portrait rights, or a labelled AI-generated text that copies protected material, remains unlawful. The label is a transparency measure, not a safe harbour. Companies that treat disclosure as a complete defence against copyright, trademark or personality-rights claims will find themselves exposed.
This distinction is worth bearing in mind because the temptation for businesses will be to label content and consider the matter closed. In practice, labelling addresses only the transparency requirement; separate legal obligations around intellectual property, data protection and consumer law continue to apply in full.
Practical consequences for business
For most retailers and service companies, the immediate compliance burden is lighter than feared. E-commerce businesses using AI to generate product descriptions, translate catalogue entries or draft promotional emails face no mandatory labeling under the current rules. The regulation does not require a general audit of all AI use within a company.
The obligations that do exist are specific and enforceable. Any firm deploying a customer-facing chatbot must ensure the interface clearly identifies itself as AI-driven. Any firm publishing synthetic images, audio or video that could be mistaken for genuine footage of real people or events must attach a visible label. Enforcement falls to national authorities, and penalties for non-compliance can be significant.
People mentioned
-
Frank Fischer
Organisations
Rödl · European Union
