AI content labelling — testing consent, deep fakes, SGI, deployer notices
Primary statement
AI content labelling per EU AI Act Article 61 (informed consent for real-world testing) + Article 50.4 (deep fake disclosure) + Article 53 (GPAI technical documentation) + MeitY ITR2026.1 (SGI identification) + Adv2024.1 (user consent and labelling) + NIST AI RMF + CERT-In. Layered labelling covers testing + outputs + GPAI.
Audit-fatigue payoff
A unified labelling programme — testing consent + deep fake disclosure + SGI labelling + deployer notices — satisfies AI labelling requirements across all 4 contributing frameworks.
Strictness matrix
Scope
Scope: deep fakes (image, audio, video) + AI-generated text + AI testing subjects. Multi-modality + multi-context.
Ceiling source: eu_ai_act:Art.50.4
Rationale: EU AI Act Art 50.4 multi-modality scope is comprehensive.
Threshold
Threshold: freely-given INFORMED consent for real-world testing prior to participation. Binary consent qualifier.
Ceiling source: eu_ai_act:Art.61
Rationale: EU AI Act Art 61 informed consent threshold is uniquely strict.
Method
Method: deep fake disclosure mechanism (from Aug 2026) + AI-generated text labelling + SGI identification per MeitY ITR2026.1 + testing consent procedure per Art 61 + deployer-side user consent per Adv2024.1.
Ceiling source: eu_ai_act:Art.50.4
Rationale: EU AI Act Art 50.4 + Art 61 + MeitY ITR2026 combined are the most prescriptive.
Frequency
Per output: continuous (automated). Testing consent: per testing session. Programme review: annual.
Ceiling source: eu_ai_act:Art.50.4
Rationale: Per-output continuous labelling is the operational floor.
Evidence
Evidence: deep fake disclosure implementation + SGI labelling + testing consent records + sample labelled outputs.
Ceiling source: eu_ai_act:Art.50.4
Rationale: EU AI Act Art 50.4 evidence is the audit-defensible anchor.
Auditor test pattern
Step 1: For deep fake / generative AI, verify Art 50.4 labelling. Step 2: For real-world testing, verify Art 61 informed consent. Step 3: Verify SGI labelling per ITR2026.1. Step 4: Inspect sample labelled outputs.
Common findings
Common findings: (1) Deep fake labelling absent — Art 50.4 prep not started; (2) Testing consent informal; (3) SGI labelling missed for significant intermediaries; (4) Labelling visible-only without machine-readable metadata.