Do you have to label AI-generated content under the EU AI Act?

The short answer

Article 50 of the EU AI Act splits the job in two. The companies that make generative AI systems (providers) must mark what the system produces in a machine-readable way, so software can detect that it is AI-generated. The businesses that use those systems to publish (deployers) must show a visible disclosure only in two cases: deepfakes, and text published to inform the public on matters of public interest. Even for that text, no label is needed if a person reviewed it and someone holds editorial responsibility for publishing it.

For most small businesses posting marketing content, that means: no legal duty to put "made with AI" on every post, but a real duty on deepfake images and video, a practical reason to keep a record of human review, and tools that mark their output for you. This page explains general rules, not legal advice for your situation.

Who has to do what

The general date comes from Article 113, which applies the Act from 2 August 2026. The later date for marking by existing systems comes from the Digital Omnibus amendment adopted in July 2026; according to Slovenia's communications regulator, it covers only Article 50(2) for systems placed on the market before 2 August 2026, and does not move the deepfake or text disclosure rules.

What this means for a marketing post

• A product photo or illustration made with AI is not a deepfake unless it resembles a real person, object, place or event and would falsely appear authentic. A generated latte on a table is usually not; a generated "photo" of a real customer, your real shop front or a real event usually is, and needs a disclosure.
• A promotional post about your opening hours or a new product is rarely "text published to inform the public on matters of public interest". A blog article on a health topic, local news or a public issue may be. If you are unsure, the review exemption is the safe route.
• An article a person reviewed and approved does not need the disclosure under 50(4), provided someone holds editorial responsibility. That is why the record of review matters: it is your evidence.

What machine-readable marking looks like

The Act does not prescribe a format. The Commission's Code of Practice on marking and labelling AI-generated content, published in final form on 10 June 2026, is a voluntary way to show compliance; signing it is optional, the obligations are not. Common techniques for images are metadata and signed manifests: the IPTC "digital source type" field with the value for trained algorithmic media, which several platforms read for their AI labels, and C2PA Content Credentials. For web text, schema.org now has a digitalSourceType property that can carry the same value in an article's structured data.

What ADpilot does

ADpilot generates images and text and publishes them for you, so it takes on the provider side and gives you the tools for the deployer side. As built today:

• Images it generates are written with IPTC digital source type metadata, the AI system used and a credit line. When your logo is added on top, the file is marked again as a composite of AI and other material, because editing the image would otherwise strip the metadata. Stock photos are left unmarked.
• Articles published to WordPress or through a webhook carry schema.org digitalSourceType in their structured data and a short HTML comment at the top of the body. Text a person edited before approval is marked as a composite rather than as purely AI-generated.
• A visible label is your choice: a switch in brand settings adds one.
• The review record. Every draft waits for approval by default. On approval ADpilot stores who approved, when, and which text fields a person edited; an approval made by an automation is recorded as not human-reviewed. The whole record exports as a CSV.

What to do now

1. Ask every AI tool you use whether, and how, it marks its output. Keep the answer.
2. Decide who in your business approves what is published, and make sure that approval is recorded with a name and a date.
3. Label any image or video that shows a real person, place or event that did not happen that way.
4. If you publish on public-interest topics such as health, finance or politics, either review every piece or label it.
5. Check the Commission's Article 50 guidance again before the end of 2026; guidelines and the Code of Practice will keep developing.

Where ADpilot shows this

Approval workflow in ADpilot shows this for your own brand.