For marketing teams, the practical question is not simply, ‘Was AI involved?’ It is, ‘What was generated or changed, how will people interpret it, and which obligations apply to this use?’ Article 50 distinguishes between technical marking by AI-system providers, disclosures for deepfakes and certain public-interest text, and notices for direct interaction with AI. [1]
What the rules distinguish
‘AI label’ can describe several different things. Treating them as interchangeable makes compliance decisions harder.
| Transparency measure | What it addresses | Why marketers should distinguish it |
|---|---|---|
| Machine-readable marking | Article 50(2) requires providers of relevant generative AI systems to make outputs machine-readable and detectable as artificially generated or manipulated, subject to specified exceptions. [1] | This is not identical to adding a visible caption to an advertisement. |
| Deepfake disclosure | Article 50(4) requires deployers to disclose AI-generated or manipulated image, audio or video content constituting a deepfake. [1] | The content and its presentation matter—not just the software used. |
| Public-interest text disclosure | Article 50(4) addresses AI-generated or manipulated text published to inform the public on matters of public interest, with a human-review/editorial-control exception that also requires editorial responsibility. [1] | Public-interest publishing needs a different assessment from routine copy editing. |
| AI-interaction notice | Article 50(1) addresses systems interacting directly with people, unless the AI interaction is obvious under the statutory test. [1] | A chatbot raises a different transparency question from a static campaign image. |
A business may occupy more than one role. Do not assume that using a third-party tool answers every question about your own responsibilities.
Does every image need labelling?
No—not every AI-generated image automatically requires a visible label under Article 50(4). The Commission expressly distinguishes the content covered by that provision from AI-generated content generally. [2]
The deepfake definition covers AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful. [2]
That makes context important. Consider these illustrative review scenarios, not automatic legal classifications:
| Marketing scenario | Question to investigate |
|---|---|
| A clearly stylised illustration | Could anyone reasonably mistake it for an authentic depiction of an existing person, object, place or event? |
| A real apartment photograph furnished using AI | Could viewers believe the generated furnishings were present when the photograph was taken? |
| A synthetic voice resembling a recognisable spokesperson | Does it appear that the person genuinely spoke the words? |
| An AI-assisted image of a product | Could the image misrepresent the actual product’s appearance, capabilities or included features? |
| An obvious fictional campaign scene | Is it evidently creative or fictional, and what disclosure remains appropriate? |
The Commission’s own examples include AI-furnished apartment photographs. The deepfake assessment is therefore not confined to celebrity impersonations or political videos. [2]
Human review has limits
A common misunderstanding is that human approval removes every AI disclosure requirement.
For public-interest text, Article 50(4) provides an exception where the content undergoes human review or editorial control and a natural or legal person holds editorial responsibility for publication. [1]
That is not a general exemption for images, audio or video constituting deepfakes. Those materials are addressed separately in the same provision. [1]
Likewise, Article 50(2)’s exception for certain standard editing or changes that do not substantially alter the input or its semantics concerns the provider-side marking obligation. It should not be treated as a blanket exemption covering every publishing decision. [1]
Recommended practice: record what the reviewer actually checked. ‘Approved’ is less informative than a record explaining the source material, changes, factual checks, disclosure decision and responsible approver.
Will labels hurt performance?
The research does not justify a universal prediction.
A 2025 advertising experiment found lower trust in the advertisement and organisation in its tested disclosure setting. [3] Another study identified competing effects: AI labels increased perceived novelty while reducing perceived authenticity. [4]
Neither finding establishes the sales impact of a particular Staimp reader’s campaign.
Recommended approach: test how to communicate clearly—not whether to hide a disclosure that is legally required. Where several compliant presentations are available, evaluate their readability, audience understanding and campaign performance.
Useful testing questions include:
- Do viewers understand which part of the asset was generated or changed?
- Does the label remain readable on mobile?
- Does the wording accurately describe the production process?
- Could the creative itself mislead people about a product, person or event?
- Does the disclosure survive the format in which the campaign will actually appear?
Make the disclosure specific
Where a disclosure is needed, use wording that accurately identifies the AI involvement.
Illustrative wording—not pre-approved legal templates:
- ‘AI-generated image.’
- ‘Background generated with AI.’
- ‘Voice generated with AI.’
- ‘Image modified with AI.’
Only use a narrower statement such as ‘background generated with AI’ if it accurately describes the extent of the AI involvement.
Article 50 requires the relevant information to be clear and distinguishable, accessible, and provided no later than the first interaction or exposure. [1] A notice buried in a general website footer should therefore not be assumed to satisfy an asset-specific disclosure requirement. [1]
The EU’s official AI icons are optional. Using one does not establish legal compliance by itself. [2]
The Commission’s implementation guidance also addresses visibility when content is downloaded or reshared. Its code-related placement measures should be distinguished from the underlying statutory requirements. [2]
Build a review process
Recommended pre-publication workflow:
- Identify the asset. Record the final version, destination and intended audience.
- Document AI involvement. Establish what was generated, edited or replaced.
- Identify your role. Determine whether the organisation is acting as a provider, deployer or in more than one capacity.
- Assess the content. Review possible deepfake characteristics, public-interest text and direct AI interaction.
- Check applicable exceptions. Record the reasoning rather than relying on an informal assumption.
- Check other requirements. Include relevant platform, advertising, privacy, intellectual-property and contractual reviews.
- Approve the final presentation. Check wording, placement, accessibility and the published rendition.
- Retain a decision record. Keep the evidence and identify who approved publication.
This is an editorial workflow recommendation, not a claim that every step is a separately prescribed Article 50 obligation.
Frequently asked questions
Is there a mandatory EU icon?
No. The Commission’s icons are optional, and using them does not by itself establish compliance. [2]
Does human review remove image disclosures?
Not automatically. The human-review/editorial-control exception in Article 50(4) concerns the specified public-interest text obligation, not a general exemption for deepfake images, audio or video. [1]
Can an agency rely on the generator?
The generator’s machine-readable marking and the deployer’s disclosure duties are distinct. A provider-side measure should not be assumed to discharge a separate deployer obligation. [1]
Should campaigns stop using AI?
That is a commercial decision, not the conclusion of Article 50. Start by assessing the particular content and obligations rather than treating all AI use as one category. [1]
Before publishing, review the asset—not just whether AI was involved. Explore Staimp.
Compliance notice
This article provides general information, not legal, regulatory or other professional advice. It reflects the sources available on the stated review date; laws, guidance and interpretations may change. Requirements depend on your role, content, use case and jurisdiction. Examples are illustrative. Automated scores or checks are not legal determinations or compliance certification. Verify current official requirements and obtain qualified advice before making decisions. No article, label, disclaimer or tool result guarantees compliance.