// AI Act · Art. 50 Protocol
AI Labelling.
Art. 50 from 2 August 2026.
From 2 August 2026, Art. 50 of the EU AI Act requires AI content to be made transparent: deepfakes and certain AI texts must be clearly disclosed as AI-generated. We make your AI content compliant — without slowing your production down.
of the EU AI Act governs the transparency obligations
binding and applicable from this date
or 3 % of worldwide turnover — fine per breach
defined very broadly — almost any photorealistic AI image
// The Obligation
What Art. 50
requires of you.
Art. 50 of the EU AI Act (Regulation (EU) 2024/1689) governs the transparency obligations for AI content. It distinguishes between providers (who build the AI system) and deployers (who use AI professionally). For companies and agencies, the deployer role is usually the relevant one.
Important: the label says nothing about truth or quality — only about the artificial origin. It protects public opinion-forming and trust in the information ecosystem.
A deepfake is defined in Art. 3(60) AI Act as AI-generated or manipulated image, audio or video content that resembles real persons, places or events and would falsely appear authentic. The wording is extremely broad — as of today almost any photorealistic AI image can fall under it.
Note: AI Pirates is an AI agency, not a law firm. This page describes the situation in plain language. For a binding legal assessment of your individual case, please coordinate with your legal department.
The 4 core obligations
// Implementation
How to label —
by format.
Images
VISIBLE ON FIRST VIEW
Videos
NOT BEHIND UI ELEMENTS
Audio
PODCASTS, VOICEBOTS, SPOTS
Text
BLOG, NEWS, PR
Visibility beats metadata: C2PA metadata is a good technical foundation but is lost on re-upload and screenshots — and does not exempt you from visible labelling.
// Exceptions
When you don't have
to (fully) label.
Art & satire
For evidently artistic, creative or satirical works the obligation is softened: a discreet notice that doesn't impair enjoyment of the work suffices. But: a soft, unclear exception — high liability risk around „is this already art?".
Human-in-the-loop
AI texts are exempt if subjected to human review / editorial control and a person bears editorial responsibility. A demonstrable final human edit usually exempts you.
Law enforcement
The disclosure obligation does not apply where use is legally permitted to detect, prevent, investigate or prosecute criminal offences. Rarely relevant for companies and agencies in practice.
// Risk & deadlines
What's at
stake.
Sanctions
8 May 2026 · Draft guidelines
The EU Commission publishes draft guidelines on Art. 50.
10 June 2026 · Code of Practice
Voluntary code published — eases the compliance proof. First signatory deadline: 22 July 2026.
2 Aug 2026 · Obligation applies
Art. 50 becomes binding. From now on fines and warnings are possible.
2 Feb 2027 · Interoperability
Deadline for interoperable watermark detectors (Code of Practice).
// How to do it right
Meet the obligation,
keep producing.
When in doubt, label
Whether and how to label is case-by-case. Given the broad deepfake definition, defensive, generous labelling is the safest route.
Tool whitelisting & policy
Only vetted AI tools with compliant watermarks / C2PA, a clear internal AI policy, clauses for external creators. We build that for you.
Secure your evidence
Document original prompts, apply a four-eyes principle before publishing. That proves the human share and keeps you covered if questioned.
// FAQ
Frequently asked
about labelling.
When does the AI labelling obligation under Art. 50 apply?
From 2 August 2026. AI providers must mark outputs in a machine-readable way; AI deployers must disclose deepfakes and certain AI-generated texts clearly and perceptibly as AI-generated.
Do AI-generated images have to be labelled?
Usually yes. The deepfake definition (Art. 3(60) AI Act) is very broad: as of today, any photorealistic AI image resembling real persons, objects or events can be a deepfake that requires labelling. When in doubt, label.
What does an AI label have to look like?
Clear, unambiguous and accessible, at the latest on first view. Recommended: a label „AI" capitalised, contrast at least 4.5:1, and platform-appropriate display — continuous for short videos, spoken notices for audio.
Do AI-generated marketing texts have to be labelled?
Mostly not. The text obligation applies primarily to text of public interest. Purely promotional text often falls outside it, and a final human review with editorial responsibility additionally exempts it.
Are C2PA metadata enough for labelling?
No. Metadata is a useful technical foundation but is easily lost (re-upload, screenshots). The AI Office acknowledges this: presence of metadata does not exempt you. Deepfakes and relevant AI texts must be labelled perceptibly for humans — visibility beats metadata.
What are the penalties for non-compliance?
Under Art. 99 AI Act up to €15M or 3 % of worldwide annual turnover (whichever is higher). Plus competition-law warnings by competitors and consumer bodies. For an individual legal assessment we recommend your legal department.
// Related AI Services
// Get ready now
Make your AI content
compliant.
Free intro call — 30 minutes. We review your AI content, the labelling risk and build your policy, tool whitelist and training. 2 August 2026 comes faster than you think.
Request a labelling audit