AI Disclosure Materiality Test
Tags: Frameworks
TL;DR
- A four-step test decides if AI involvement must be disclosed to avoid misleading consumers.
- Labels are required when AI materially affects authenticity, identity, representation, or an important decision.
Why it matters for HK marketers: It reduces both over-labeling noise and under-labeling risk that can trigger complaints or enforcement.
The four-step decision
- AI involvement: Was AI used to create or materially alter the asset/interaction? If no, no AI-specific disclosure.
- Consumer understanding: Could AI involvement change a reasonable consumer’s view of authenticity/identity/representation/decision basis? If no, retain internal record and provenance where feasible.
- Risk of deception/harm: Would omission of a label create meaningful risk? If yes, add clear, proximate disclosure and provenance controls before release.
- Lawfulness/accuracy: If the content is unlawful, inaccurate, unfair, or unauthorised—do not publish; labeling is not a cure.
Execution details
- Placement: Match notice to medium (on-screen for video, audible for audio, opening message for agents).
- Evidence: Keep the assessment, approved wording/placement, live screenshots/recordings, provenance metadata, rights/consent records, and platform requirements.
- Escalation: Deepfakes, cloned voices, synthetic humans in primary roles, fabricated events/locations, political/public-interest content, and vulnerable audiences.
Example
A synthetic influencer describing a hotel stay requires a prominent disclosure and embedded provenance; a stylised skyline illustration likely does not, but the creation record is retained.
So what for marketers
Bake this test into pre-launch QA and creator/influencer briefs; no asset should go live without a recorded decision and proof of label persistence where applied.
Sources:- (June 14) Responsible AI in Digital Marketing Playbook - Executive Summary & Checklists.pdf
← Back to Knowledge Base