Encounters with AI
Synthetic content
The judgement
Synthetic-content transparency has two layers. Providers must build marking into many AI outputs, while professional deployers publishing particular content must give people a perceptible disclosure. These are duties on those organisations, not a guarantee that every viewer can personally inspect a hidden technical mark.
A deepfake is an AI-generated or manipulated image, audio or video resembling existing people, objects, places, entities or events and falsely appearing authentic or truthful. That definition has applied since 2 February 2025.1
| Situation | Applicable rule |
|---|---|
| A system generates synthetic audio, images, video or text | Under Article 50(2), the provider must make outputs machine-readably marked and detectable so far as technically feasible. The duty excludes qualifying standard editing and specified legally authorised criminal-law uses. National market-surveillance authorities, the limited-scope AI Office or the EDPS have enforced it since 2 August 2026.2 |
| The generative system entered the market before 2 August 2026 | Article 111 gives its provider until 2 December 2026 to meet the marking duty. The same authorities enforce it from that later date for those systems.2 |
| A professional deployer publishes a deepfake | Article 50(4) requires that deployer to disclose the artificial generation or manipulation. The relevant AI Act authorities have enforced the duty since 2 August 2026.2 |
| A professional deployer publishes synthetic text about a public-interest matter | Article 50(4) requires disclosure, enforced by the same authorities since 2 August 2026. The duty does not apply when substantive human review or editorial control occurred and someone holds editorial responsibility.2 |
| A deepfake forms part of an evidently artistic, fictional, creative or satirical work | Article 50(4) still requires the professional deployer to disclose it, under the same enforcement and start date, but permits an appropriate presentation that does not hamper enjoyment.2 |
Article 50(5) requires deployers making these disclosures to provide a clear, distinguishable and accessible label by first exposure. The relevant AI Act authorities have enforced that deployer duty since 2 August 2026. Commission guidance says the disclosure should be visible or audible and cannot rely only on the provider’s hidden mark.3
A person who suspects a marking or labelling breach can use the Article 85 complaint route to the relevant market-surveillance authority. National procedures differ, and no single EU-wide form or consumer tool for checking individual machine-readable marks was identified.2
The high-risk postponements to 2027 and 2028 do not defer these duties. The narrow exception is the transition ending in December 2026 for older generative systems.2
Boundary. A missing label does not by itself prove that content is false. Standard editing that does not substantially alter an input or its meaning can also fall outside the provider marking duty.
Quellen
Quizze
A professional organisation uses a generative system to publish a realistic AI-manipulated video of an existing person. Which obligation belongs to the organisation as publisher, rather than to the system provider?
- Give viewers a clear disclosure by their first exposure
- Make generated outputs machine-readably marked and detectable
- Build an interoperable technical marking solution into the system
The professional deployer owes the audience-facing disclosure. Machine-readable and detectable marking of generated outputs is the system provider’s separate technical duty.
A professional publisher cannot satisfy its audience-facing deepfake duty solely through ____.
- a hidden machine-readable mark
- a visible first-exposure label
- an audible first-exposure notice
The provider’s technical mark and the deployer’s perceptible disclosure are separate layers. Viewers must receive a visible or audible disclosure where required.
Kommentare
Noch keine Kommentare. Fang das Gespräch an.