EU AI Act — Article 50 Transparency Evidence
Article 50 transparency obligations apply from 2 August 2026 to most organisations using generative AI — not only high-risk systems. UI labels and policy PDFs describe what you intend; signed events prove disclosure happened when a regulator asks years later.
Labels are not evidence
Article 50 requires providers and deployers to inform users when they interact with AI or encounter AI-generated content. That work is partly UX design and partly technical marking (Article 50(2)). Labels and watermarks fulfill the transparency duty — they do not by themselves create a durable, checkable audit trail. Sigigo can witness disclosures as signed events so you can show when and how the notice ran, years later.
Emit ai_act.transparency_disclosure at the disclosure boundary — when the user first sees the notice, when synthetic output is published, or when biometric categorisation begins. Link disclosures to the same session_id as inference events so an evidence pack reconstructs the full chain.
Four situations → what to log
Map each Article 50 situation to a signed disclosure event. Sigigo witnesses that the obligation was executed — not how you designed the label (see FLI for legal detail and Code of Practice marking standards).
| Article 50 | User-facing duty (summary) | Sigigo event | Key metadata |
|---|---|---|---|
| 50(1) — AI interaction | Users know they interact with AI (chatbots, agents, automated phone) | ai_act.transparency_disclosure | disclosure_type: chatbot, channel, locale, first_interaction: true |
| 50(2) — Synthetic content | Machine-readable marking on AI-generated audio, image, video, text | ai_act.transparency_disclosure + ai_act.output_generated | marking_applied: true, output_hash, marking_standard (e.g. C2PA ref) |
| 50(3) — Emotion / biometric categorisation | Deployer informs persons exposed (screen Art. 5 prohibitions first) | ai_act.transparency_disclosure | disclosure_type: biometric_notice, exposure_context |
| 50(4) — Deepfakes & public-interest text | Deployer discloses synthetic media or AI-generated public-interest text | ai_act.transparency_disclosure | disclosure_type: deepfake | synthetic_media | public_interest_text |
Omnibus grace: generative AI on market before 2 August 2026 may defer machine-readable marking under 50(2) until 2 December 2026 — disclosure events should still be logged from day one.
Engineering checklist
- Emit
ai_act.transparency_disclosureat the first interaction or exposure — not only in terms of service - Include channel (web, mobile, API, voice), locale, and disclosure copy hash — not necessarily full marketing text in the canonical payload
- Link disclosure events to
ai_act.inference_started/ai_act.output_generatedvia sharedsession_id - For 50(2), log
marking_appliedat the generation boundary — Sigigo does not render EU labels or watermarks - Retain disclosure events for at least six months (Art. 26(6) deployer retention where applicable)
- Export a session slice in the evidence pack for surveillance or contractual audit
Typical compliance gap
| Approach | Why it falls short |
|---|---|
| UI label only | No signed record; design can change without audit trail |
| Policy PDF / intranet page | Proves intent, not per-session execution |
| Screenshot of chatbot disclaimer | Mutable; no chain to inference events |
| Vendor marking tool only | Proves mark exists, not that user saw disclosure at first interaction |
From Article 50 to Article 12
Article 50 reaches every generative-AI deployer; Article 12 automatic logging applies to high-risk Annex III systems by December 2027. Teams that instrument transparency events now extend the same ai_act.* pipeline to inference, human oversight, and evidence packs — see the Article 12 guide.
FAQ
- Does Article 50 apply to high-risk AI only?
- No. Transparency obligations apply across risk tiers — including chatbots and synthetic content that are not Annex III high-risk. High-risk systems have additional Article 12 logging duties.
- Should Sigigo build watermarking or EU AI labels?
- No. Machine-readable marking under 50(2) is a separate product lane (C2PA, metadata standards, Code of Practice). Sigigo witnesses marking_applied and disclosure events at your application boundary.
- What if we already show a chatbot disclaimer?
- Good UX is necessary but not sufficient for later defensibility. Log ai_act.transparency_disclosure when the user first encounters the notice so you can export a signed record years later — that integrity layer is Sigigo, not an Article 50 mandate.
- Where do I get legal classification help?
- Use the FLI Compliance Checker for role and risk tier — then map technical evidence using this guide and the Article 12 field map.
Read the evidence-angle blog post: Article 50 Transparency: Labels Aren't Evidence. Map disclosures to ai_act.* events — get in touch.
This guide supports qualification and architecture planning — it is not legal advice. Work with your legal and compliance teams on classification, DPAs, and deployment approvals. See also our Regulatory & Compliance page.