Regulation (EU) 2024/1689 · Art. 50(2)
Is my AI system exempt from Article 50(2)?
Most systems are in scope. The three narrow carve-outs, what keeps you in scope, and the exact wording.
Updated 16 October 2026. From Art. 50 of Regulation (EU) 2024/1689, the Commission Guidelines and the Code of Practice (June 2026). A vendor publication by Verda Inc. Not legal advice. Last change: Article 50 applies from today, 2 August 2026.
Exemption test for Article 50(2)
Narrow ways out
Most systems are in scope. Most voice-agent calls are recorded by at least one party, so the real-time carve-out rarely applies.
−You may be outside if
- Assistive standard editing that does not substantially alter the input
- Output authorised by law for criminal-offence work
- Real-time, unrecorded and not feasible to mark (Guidelines)
+You are in scope if
- It generates new content or changes its meaning
- Real-time output is recorded, stored, forwarded or replayed
- You provide a general-purpose model others use
Exact wording and sources
You may be outside 50(2) if: the system performs an assistive function for standard editing (colour correction, noise reduction, crop) and does not substantially alter the input or its semantics; the output is authorised by law for detecting, preventing, investigating or prosecuting criminal offences; or, under the Guidelines, content is generated in real time, is not recorded, and marking is not technically feasible.
You are in scope if: the system generates new audio, image, video or text, or changes the meaning of existing content; real-time output is recorded, stored, forwarded or replayed anywhere in the chain; or you are a general-purpose model provider whose outputs are used by others.
Sources: Art. 50(2) second subparagraph; Commission Guidelines on Article 50. The real-time carve-out is narrow: most voice-agent calls are recorded by at least one party. Get a use-case review before relying on it.
FAQ
Common questions
Which AI systems are exempt from the Article 50(2) marking duty?
Three narrow cases: a system performing an assistive function for standard editing (colour correction, noise reduction, crop) that does not substantially alter the input or its semantics; output authorised by law for detecting, preventing, investigating or prosecuting criminal offences; and, under the Commission Guidelines, content generated in real time that is not recorded, where marking is not technically feasible.
Is a real-time voice agent call exempt from Article 50(2)?
Only if the audio is not recorded and marking is not technically feasible. Real-time output that is recorded, stored, forwarded or replayed anywhere in the chain is in scope, and most voice-agent calls are recorded by at least one party. Get a use-case review before relying on the carve-out.
Do general-purpose AI model providers have to mark outputs under Article 50(2)?
Yes. A general-purpose model provider whose outputs are used by others is in scope of Article 50(2).
If my system is outside Article 50(2), do I still have Article 50 duties?
They may still apply. Disclosure duties under Article 50(1) and 50(4), AI interaction notices and deep-fake disclosure, are separate from the 50(2) marking duty.
More in the playbook
Other sections
Eight steps for providers
Mark at generation, attach a C2PA manifest, offer free detection, test, document. A checklist you can tick, with the detail for each step.
Open the checklist → Vendors comparedWhat each vendor says about itself
Verda, SynthID, Adobe Content Authenticity and Meta AudioSeal / Video Seal on eight criteria, from each company's public site.
Compare vendors → Deployer checklist · Art. 50(4)What the person sees
Deep-fake and public-interest text disclosure at first exposure. A watermark does not discharge these duties.
Deployer checklist → TimelineDeadlines
New systems mark from 2 August 2026, systems already on the market from 2 December 2026, interoperable detection by 2 February 2027.
Full timeline →