AI labeling since August 2, 2026

The deadline is set

The “Digital Omnibus” – Regulation (EU) 2026/1744, in force since July 2026 – postpones the obligations for high-risk AI until December 2027 and August 2028, respectively. The transparency obligations under Article 50 of the AI Act are not affected. They have applied since 2 August 2026.

At the same time, there is no blanket labelling requirement for everything created with AI. Two questions are decisive: what role are you acting in, and what kind of content is involved?

Providers or deployers—the roles are clearly distinct

The Regulation distinguishes between providers, who develop an AI system and place it on the market, and deployers, who use a finished system in their own business operations. Both have obligations, but they differ. Most companies fall into the deployer role; however, anyone who significantly modifies a system they have purchased and operates it under their own name may themselves become a provider.

Four everyday business scenarios

1. Chatbots and voice assistants

Anyone chatting about services, availability or technical questions must be able to tell that they are interacting with an AI system—unless this is already obvious. The same applies to voice assistants on customer service hotlines and to AI avatars.

This obligation primarily applies to the provider, who must design the system accordingly. As a deployer, you should nevertheless check that the notice is actually visible in your implementation. Incidentally, the courts have already made clear, independently of the AI Act, that companies may be liable for false statements made by their chatbots.

2. AI-generated images, videos and audio content

This is where most companies face the greatest practical risk, because it affects all marketing materials.

Labelling is required for content that the Regulation defines as a deepfake: content generated or altered by AI that resembles real people, objects, places or events and could be wrongly believed to be authentic. Typical examples include:

• A generated image of company premises, a workshop or facilities for a corporate brochure
• A photorealistic image of a product or vehicle showing a state that does not actually exist
• An AI-retouched existing photo in which damage, signage or the surroundings have been changed
• Visualisations of trade fair appearances or reference projects that could pass for photographs
• Synthetic voices in corporate image or training videos

This does not include clearly artificial illustrations, technical diagrams or obviously fictional depictions. Purely technical edits such as noise reduction, colour correction or cropping are also unproblematic. Separately, platforms such as LinkedIn may enforce their own, stricter rules.

One point that is often overlooked: with depictions of products and vehicles, there is a second consideration in addition to the AI Act. An enhanced image can have consequences under competition and consumer law, regardless of the labelling question.

3. Published AI-generated texts

If an AI-generated text is published to inform the public about matters of public interest, the use of AI must be disclosed. This may be relevant, for example, to press releases, industry articles or statements on infrastructure and regulatory issues.

The obligation does not apply if the content has undergone human review or editorial oversight and a natural or legal person bears editorial responsibility. Important: the Commission requires substantive review involving professional judgement. A spellcheck or formal review is not enough.

4. Quotes, documentation and internal materials

If AI is used only to structure, draft or edit content and the result is professionally reviewed, this does not create a visible labelling obligation. This applies equally to minutes captured through speech dictation, AI-assisted quotes and internal presentations. The purely personal use of AI-generated content is also generally not subject to labelling requirements.

What about existing content?

According to the European Commission’s guidelines of 20 July 2026, content generated and published before 2 August 2026 does not need to be labelled retrospectively.

There are two caveats. First, guidelines reflect the Commission’s interpretation; they are not binding law. Ultimately, only the European Court of Justice could provide a binding clarification. In practice, however, they serve as a reference for enforcement by supervisory authorities.

Second, the issue is only temporarily on hold: republishing or substantially revising existing content triggers the obligation. For companies with rotating campaign visuals, reused trade fair materials or an upcoming website relaunch, this is the more practically relevant point.

A second deadline is still to come

In addition to visible labelling, the Regulation requires providers of generative AI systems to use machine-readable markings—for example, watermarks or metadata. For systems already on the market before 2 August 2026, the transition period for this requirement runs until 2 December 2026. This deadline applies exclusively to providers. There is no grace period for deployer obligations—the chatbot notice, deepfake labelling and text disclosure.

What about the high-risk rules?

The postponed obligations concern AI systems used in sensitive areas, including recruitment and employee performance evaluation, creditworthiness assessments, critical infrastructure—including the transport sector—and AI used as a safety component in regulated products. Anyone using or planning to use AI in these areas should assess its classification early, even though the obligations will not take effect until December 2027 and August 2028, respectively.

Based on the current understanding, the high-risk classification does not apply to typical commercial and documentation tasks such as document research, speech dictation and preparing quotes.

Four steps

1. Create an AI inventory. Which systems are in use, who uses them, and what content do they generate?
2. Clarify your role. Provider or deployer—take a close look at systems that have been substantially modified.
3. Review your image library. Which published visuals are photorealistic and show real people, places or objects? This list is usually shorter than you might fear.
4. Assign responsibilities. Who reviews AI-generated texts, who takes responsibility, and where is this documented? Agency and freelancer contracts should also be reviewed.

Three questions to ask when in doubt

• What exactly did the AI create or change?
• Could someone mistake the content for something authentic?
• Has a human reviewed the content, and is someone taking responsibility for it?

Conclusion

Article 50 requires less than the current debate might suggest, and more than the discussion about postponements might lead you to believe. For most companies, the effort involved is manageable: take stock, set clear rules and define a review process.

The harder part is translating this into day-to-day work—who in marketing decides whether an image looks realistic, and who in sales knows when an AI-generated text must be disclosed? These are precisely the questions we address through practical business scenarios in NRailMind’s AI training courses.

Sources

• Regulation (EU) 2024/1689 (AI Act), in particular Art. 6, Art. 25, Art. 50, Art. 99(4)
• Regulation (EU) 2026/1744 (Digital Omnibus on AI)
• European Commission Guidelines on the transparency obligations under Art. 50 of the AI Act, C(2026) 5054 final, published on 20 July 2026 (press release IP/26/1653)
• Code of Practice on Transparency of AI-generated Content (voluntary)

Note: This article is intended solely to provide general information about the legal position and does not constitute legal advice within the meaning of the German Legal Services Act (Rechtsdienstleistungsgesetz). The requirements described are general in nature and do not replace an assessment of a specific case. Some of the interpretations set out here—in particular regarding the treatment of content that has already been published—are based on European Commission guidelines that are not legally binding and may be assessed differently by courts. As of: August 2026. No liability is accepted for the timeliness, accuracy or completeness of this information.

This article was created with AI assistance and editorially reviewed.