"Is this AI?" should never be a question your users have to ask. This playbook designs how the product discloses AI involvement — labels, identity, timing — plus consent moments that are real choices, and data-provenance display showing what an output was built from, aligned with EU AI Act and DPDP realities.
Workflow area: Responsible AI · Tier: Advanced · Hands-on time: ~40 minutes
Disclosure design sits at an awkward intersection: regulation increasingly mandates it (the EU AI Act requires that people interacting with an AI system be informed, and that certain AI-generated content be marked; India's DPDP Act demands notice and consent in plain language for personal-data processing), users deserve it, and product teams fear it kills conversion. The fear is mostly wrong — but only when disclosure is designed rather than bolted on. A tiny gray "AI-generated" footnote is compliance theater; a well-placed provenance panel is a trust feature users come to rely on. This playbook designs the three layers as UX: labels people notice, consent people understand, provenance people can check.
Map every surface where AI involvement could be ambiguous. Walk the product and list each point where a reasonable person might not know AI was involved: conversational surfaces (does the user know it's a bot?), generated content shown to third parties, AI-assisted human work (the support reply a human edited), recommendations and rankings, and synthetic media of any kind. For each, record audience, ambiguity level, and consequence of the misunderstanding. This map is the disclosure spec's scope — and the artifact counsel will actually engage with.
Design disclosure at three timing tiers. Match the mechanism to the ambiguity: identity disclosure — at first contact, a conversational surface says what it is ("You're chatting with our AI assistant — a human is one tap away") and never impersonates a person: no human name posing as staff, no fake typing indicators; content labels — AI-generated artifacts carry a label that travels with the artifact (in the email footer, the document header, the image metadata), not just the generating UI; process disclosure — for AI-assisted human work, disclose the blend where it changes the reader's interpretation ("Drafted with AI, reviewed by Dr. Rao"). Test every label with the squint test: visible at reading distance, in the artifact's real context, without hunting.
Turn consent moments into real choices. A consent dialog is real when it passes four checks: specific (one decision, not five bundled — memory storage and feedback signals are separate consents), timed (asked when the feature becomes relevant, not stacked at onboarding before anything means anything), symmetric (declining is one tap, same size, no guilt copy, and the product remains usable — DPDP's free-consent standard makes coerced consent no consent), and revocable (withdrawal lives where it was granted, and works — pb-44's deletion honesty applies). Draft each consent string with AI, then audit it:
You are a consent-UX reviewer applying GDPR/DPDP plain-language
standards and dark-pattern rules.
Here are my consent moments: [PASTE: trigger, string, options,
what happens on accept/decline]
For each:
1. Rewrite the string at an 8th-grade reading level, under 40
words, naming: what data, for what purpose, revocable where.
2. Flag bundling (multiple decisions in one ask), asymmetric
friction (accept easier than decline), guilt copy, or
pre-checked defaults.
3. Say what a user who read only the BOLD text would believe —
and whether that belief is accurate.
Do not soften findings.
The bold-text test in point 3 is the one that catches most theater: if skimming produces a false belief, the consent fails regardless of the fine print.
Design the provenance panel. For substantive outputs (answers, reports, recommendations), users should be able to open "How this was made": which sources fed it (with dates — pb-48's staleness defense), what personal context shaped it (linking to pb-44's memory cards), what the model contributed versus retrieved, and known limits ("sources checked through May 2026"). Default collapsed behind a consistent affordance — a small provenance chip on the output — expanding to the panel. One design system component, reused on every AI surface, so users learn one gesture for "check this."
Pressure-test against the strictest reader. Run the full map — labels, consent flows, provenance — through three lenses: a regulator (does every AI Act-relevant surface disclose interaction and mark synthetic content? is DPDP notice specific and revocable?), a journalist (which screenshot makes the most damning story about hidden AI or coerced consent?), and a first-time user (can they say, after five minutes, what's AI, what's remembered, and what's shared?). Fix what fails any lens; then hand the package — map, flows, strings — to counsel for the legal pass.
Disclosure and consent UX implements legal obligations — the EU AI Act's transparency articles and DPDP's notice-and-consent rules carry penalties, so treat this playbook's output as the design input to a counsel review, never as the review itself.
Disclosure done well is a trust feature, not a tax — say what's AI at the moment it matters, ask for one real choice at a time with decline as easy as accept, and give every substantive output a provenance panel one tap away.