Purpose
Art. 50 AI Act (Regulation (EU) 2024/1689) imposes on providers and deployers of certain AI systems transparency and labelling obligations, in order to limit deception, disinformation and manipulation in the information ecosystem. The obligations apply from 2 August 2026 and also cover systems that are not classified as high-risk. This skill assigns the relevant transparency obligation to a system and reviews the type of labelling.
⚠️ Currency (as of 2026-07): Art. 50 applies from 02.08.2026 — the obligations are NOT postponed.
The Digital Omnibus on AI (European Parliament 16.06.2026, Council 29.06.2026) has expressly exempted the transparency obligations of Art. 50 AI Act from the postponement. They continue to apply unchanged from 02.08.2026, as originally envisaged.
This is the point of the AI Act most frequently overlooked in practice. Only the high-riskobligations were postponed — Annex III to 02.12.2027, Annex I to 02.08.2028. In advisory practice, this postponement is widely read as a blanket deferral of the AI Act. It is not: Art. 50 covers chatbots, image generators and deepfake applications irrespective of any high-risk classification and irrespective of any postponement.
Regulatory scope Applicability Prohibitions Art. 5, AI literacy Art. 4 02.02.2025 (applies) GPAI Art. 51–55 02.08.2025 (applies) — not postponed Transparency Art. 50 02.08.2026 — remains, not postponed High-risk Annex III (Art. 6(2)) 02.12.2027 (postponed) High-risk Annex I (Art. 6(1)) 02.08.2028 (postponed) New prohibitions (NCII, CSAM) 02.12.2026 National competence in Germany: The market surveillance authority is the Federal Network Agency (residual competence under the KI-MIG), not the data protection supervisory authority or the BfDI. The KI-MIG passed the Bundestag on 11.06.2026 and the Bundesrat on 10.07.2026 , but as of 21.07.2026 had not yet been promulgated in the BGBl — a BGBl. citation therefore does not exist and must not be cited.
[unverifiziert - prüfen]Official Journal citation of the amending regulation
[unverifiziert - prüfen].
Inputs
- Function of the system (interaction with humans, content generation, emotion recognition)
- Role of the client (provider / deployer — Art. 3 AI Act)
- Type of generated content (image, audio, video, text; synthetic or deepfake)
- Deployment context (customer dialogue, editorial content, art/satire)
- Existing labeling/notices
Sub-agent architecture
Three conceptual roles share the review. An obligation assigner assigns the function to one of the four pillars of Art. 50 of the AI Act (interaction disclosure, marking of synthetic content, emotion recognition, deepfake disclosure) and determines whether a provider or deployer obligation applies. A labeling designer determines how the labeling must be done technically (machine-readable) and vis-à-vis humans and which exceptions apply. A sanction assessor quantifies the fine risk under Art. 99 of the AI Act. The obligation assigner hands over each affirmed obligation to the labeling designer.
Process
1. Disclosure of AI interaction (Art. 50(1) of the AI Act)
Providers must design AI systems intended for direct interaction with natural persons (e.g. chatbots) in such a way that the affected person is informed that they are interacting with an AI system — unless this is obvious from the circumstances.
2. Marking of synthetic content (Art. 50(2) of the AI Act)
Providers of systems that generate or manipulate synthetic image, audio, video or text content must mark the output in a machine-readable format and make it recognizable as artificially generated/manipulated (e.g. watermarks, metadata).
3. Emotion recognition and biometric categorization (Art. 50(3) of the AI Act)
Deployers of emotion recognition or biometric categorization systems must inform the affected persons about the operation (interlocking with GDPR obligations).
4. Deepfake disclosure (Art. 50(4) of the AI Act)
Deployerswho generate or manipulate deepfakes (image, audio, video) with an AI system must disclose that the content was artificially generated/manipulated. For obviously artistic, creative or satirical works, a attenuated disclosure applies that does not impair the presentation. A disclosure obligation also applies to AI-generated/edited texts intended to inform the public about matters of public interest.
5. Form and timing
Information must be provided clearly and distinguishably at the latest at the time of the first interaction/exposure , in an accessible manner. Markings must be effective, interoperable, robust and as technically feasible as possible.
6. Sanctions (Art. 99 of the AI Act)
Violations of transparency obligations under Art. 50 of the AI Act can be penalized with fines of up to EUR 15 million or 3% of worldwide annual turnover .
Risks
- Forgotten labeling — synthetic content without machine-readable marking violates Art. 50(2) AI Act.
- Deepfake not disclosed — failure to disclose under Art. 50(4) AI Act; the art/satire exception is narrow and does not fully exempt.
- Chatbot without notice — failure to disclose AI interaction under Art. 50(1) AI Act.
- Postponement of high-risk obligations misunderstood as all-clear for Art. 50 — BLOCKER and the most common mistake in practice. The Digital Omnibus on AI has only postponed the high-risk obligations (Annex III → 02.12.2027, Annex I → 02.08.2028) and expressly excluded Art. 50. The transparency obligations apply from 02.08.2026. Clients who infer a deferral from the headline "AI Act postponed" run into applicability without labeling.
- Art. 50 coupled to high-risk classification — the transparency obligations apply independently of Art. 6 and Annex III. A chatbot or image generator without a high-risk connection is fully covered; the assessment "not high-risk, so nothing to do" is wrong.
- Implementation lead time underestimated — watermarks, C2PA/metadata pipelines, UI notices and editorial approval processes are technical projects. Anyone who starts them only at the deadline will not be compliant on 02.08.2026.
- National competence wrongly addressed — the market surveillance authority is the Bundesnetzagentur (residual competence, KI-MIG), not the data protection supervisory authority or the BfDI. As of 21.07.2026, the KI-MIG had not yet been promulgated in the BGBl.; a BGBl. reference must not be provided.
[unverifiziert - prüfen] - Fine risk — up to EUR 15 million / 3% under Art. 99 AI Act; enforceable from 02.08.2026.
Output format
KI-TRANSPARENZPFLICHTEN — PRÜFUNG — <System> — <Datum>
I. Funktion [Interaktion / Inhaltserzeugung / Emotion]
II. Rolle (Art. 3 KI-VO) [Anbieter / Betreiber]
III. Einschlägige Pflicht (Art. 50 KI-VO)
Abs. 1 KI-Interaktion [betroffen? <…>]
Abs. 2 synthetische Inhalte [betroffen? Markierung <…>]
Abs. 3 Emotionserkennung [betroffen? <…>]
Abs. 4 Deepfake [betroffen? Offenlegung <…>]
IV. Form / Zeitpunkt der Kennzeichnung [<…>]
V. Ausnahme (Kunst/Satire/offensichtlich) [greift? <…>]
VI. Bußgeldrisiko (Art. 99 KI-VO) [bis 15 Mio. EUR / 3 %]
Empfehlung: <…>
Nächste Schritte: <…>
Sources
Statute
- Regulation (EU) 2024/1689 (AI Act / EU AI Act) — full text
- Art. 50 AI Act (transparency obligations), Art. 99 AI Act (sanctions), Art. 113 AI Act (start of application)
- Digital Omnibus on AI — amending regulation to the AI Act (EP 16.06.2026, Council 29.06.2026); Art. 50 excluded from the postponement, application remains 02.08.2026. Official Journal reference
[unverifiziert - prüfen] - KI-MIG (AI Market Surveillance and Innovation Promotion Act) — Bundestag 11.06.2026, Bundesrat 10.07.2026, as of 21.07.2026 not promulgated in the BGBl.; market surveillance by the BNetzA
[unverifiziert - prüfen]
Guidelines
- European Commission, Guidelines on transparency obligations under Art. 50 AI Act (draft)
Secondary literature
- Hilgendorf/Roth-Isigkeit, KI-VO, 1st ed. 2024
- BeckOK KI-VO (online)