On August 2, 2026, new European Union transparency obligations began applying to certain artificial intelligence systems and AI-generated content. This change quickly produced an oversimplified conclusion: every AI-assisted text, image or audio file now needs a label.
The actual rule is more specific. It distinguishes between technology providers, professional users, content types and publication purposes. For an organization, the challenge is not to add “made with AI” to everything. It is to recognize situations in which people could be misled or may not realize they are interacting with a machine.
A sound transparency policy does not list every tool used. It discloses AI involvement when that information could change the audience’s interpretation, trust or decision.
What changed on August 2
Article 50 of the European Union Artificial Intelligence Act establishes several distinct obligations:
- Systems interacting with people: users must be informed that they are interacting with AI unless this is obvious to a reasonably well-informed person.
- Providers of generative systems: they must implement machine-readable marking so outputs can be detected as artificial or manipulated, within the regulation’s technical limits.
- Deepfakes: professional users generating or manipulating realistic images, audio or video must disclose their artificial origin.
- Public-interest text: disclosure may be required when AI-generated or manipulated text is published to inform the public, subject to specific editorial-review conditions.
- Emotion recognition and biometric categorization: exposed individuals must be informed in covered cases.
The European Commission guidelines published on July 20, 2026 clarify the scope of these requirements and support consistent implementation. The legal duties come from the regulation; the guidelines assist with their interpretation.
Four questions for deciding whether to disclose
1. Is someone interacting directly with a system?
A customer-service assistant, automated voice or conversational character should identify itself as AI when this is not already obvious. The notice should appear at the first interaction rather than being buried in terms and conditions.
2. Could the content be mistaken for an authentic event?
Cloned voices, altered photographs, videos showing someone saying words they never said and convincing representations of nonexistent events require particular attention. The concern is not only how much content was generated, but whether it can appear real.
3. Does the text inform the public about a matter of public interest?
The rules do not treat an internal draft, a product description and a publication about health, elections, security or economic affairs in the same way. Purpose and context matter.
4. Was there meaningful human review and editorial responsibility?
For certain public-interest text, the regulation provides an exception when the content has undergone human review or editorial control and a natural or legal person assumes editorial responsibility. Merely opening a document or changing one word is not a reliable governance process. An organization should be able to identify who verified the content and under which criteria.
A generic disclaimer at the bottom is not enough
Information must be clear, distinguishable, timely and accessible. In practice, the audience should receive it before treating synthetic content as authentic.
The Commission has published a voluntary Code of Practice on marking and labeling AI-generated content. Signing the code is voluntary, but Article 50 obligations are not voluntary for covered organizations.
The EU has also released optional icons for AI-generated and AI-modified content. Using an icon does not establish compliance by itself. The Commission recommends plain-language labels, accessible implementation and, where appropriate, disclosures that remain visible when content is downloaded or shared.
A useful label can answer three questions:
- Was the content entirely generated or only modified?
- Which element was altered: image, voice, video or text?
- Did a person review and take responsibility for it before publication?
Specific statements such as “voice generated with artificial intelligence” or “original image modified with AI” generally communicate more than an isolated symbol or vague disclaimer.
What organizations outside Europe should review
Being located outside the European Union does not always remove the need to assess the regulation. Its territorial scope can include providers and deployers established in other countries when an AI system’s output is used within the Union.
This does not mean that every globally accessible post is automatically covered. It does mean that organizations with European customers, users, campaigns or digital products should not dismiss the issue based only on their registered address.
A consistent policy can also prevent contradictory experiences: a chatbot identified in one market but concealed in another, a synthetic image labeled on social media but not on the company website, or different standards across marketing, communications and customer support.
A six-step review plan
- Inventory AI uses: include chatbots, voice generators, image editing, synthetic video, editorial automation and externally produced content.
- Separate organizational roles: determine when the company provides a system, integrates one or simply uses its output.
- Classify the content: distinguish minor assistance, full generation, substantial manipulation, deepfakes and public-interest text.
- Assign responsibility: decide who verifies facts, approves publication, stores evidence and selects the disclosure.
- Design labels for each channel: prepare accessible formats for websites, social platforms, audio, video, applications and downloadable files.
- Review regularly: check whether tools preserve metadata, platforms remove labels or official guidance has changed.
The inventory should also cover older content that remains in circulation, reusable templates and materials created by agencies or external contributors.
Conclusion: disclose where AI changes perception
The rules applying from August 2026 do not turn every AI-assisted correction into content that must be labeled. They focus on interactions and media where artificial origin may be material to understanding what a person is seeing, hearing or reading.
The actionable next step is to create an AI-use register and assess each case according to four variables: interaction, appearance of authenticity, public purpose and human control. Organizations facing uncertainty about territorial scope or an exception should seek qualified legal advice. When the uncertainty is primarily communicational, specific, visible and understandable disclosure is usually the stronger choice.
