AI Trust Builders — honesty & privacy UX

Trust builders make the system legible: what is AI, what left the machine, what was retained, and how to refuse. Complementary to governors (control of actions) and untrusted framing (control of prompt injection).

Pattern catalog: references/patterns.md. Adapted from tommyjepsen/awesome-ux-skills ai-trust-builders.

When to use

• Cloud providers, voice enrollment, camera, transcript retention, memory, settings / privacy copy.
• Auditing whether a feature is honest about where data goes.
• Not for: scanner/framing implementation, cost metering polarity alone, or HITL confirm shapes (ai-governors).

Seam → pattern map

Seam	Prefer	Notes
Cloud model / failover	Disclosure + Caveat + Footprints	Say when persona/recall leaves the box
Transcript / dreaming to cloud	Disclosure + Consent + Ownership + Footprints	Privacy event, not a footnote
Voice mic / enroll	Consent (+ Disclosure)	Recording indicator; gate mode ≠ consent
Camera / vision	Consent + Disclosure + Caveat	No silent always-on claim
Memory / durable facts	Ownership + Footprints + Caveat	Owner can see/delete; sources labeled
Capability / help copy	Disclosure (honest cannot-do)	False capability = bug
Chat / history	Footprints + Caveat	Name model/provider when non-obvious
Generated media	Watermark / attribution	Don't pretend human-authored

Seven patterns (short)

1. Caveat — decision-time "may be wrong"; never the only safety net.
2. Consent — opt-in for record/enroll/train; separate those acts.
3. Data Ownership — retention vs training; default privacy-protective.
4. Disclosure — label AI actor and cloud vs local.
5. Footprints — inspectable trail of what was sent/stored/applied.
6. Incognito / memory-off — session without durable write when stakes are high.
7. Watermark — mark AI-generated media when mixed with human content.

Honesty rules

• A false capability line is a bug — copy must match mechanism and flags.
• Merged ≠ live — disclosure about a path is inert until deployed.
• Degrade path must not drop disclosure.
• Caveats habituate; pair with Footprints/Citations for high-stakes acts.
• Consent buried in env comments is not consent UX.

Design checklist

• Owner can answer: local or cloud? retained? used for training?
• Voice/camera active state is perceptible without reading logs.
• Cloud paths name the provider class (not only "AI").
• Unavailable/degraded search shows notice, not empty success.
• Settings expose ownership controls or an explicit "we don't train on you".

Output format

1. Trust risk — what could be misunderstood.
2. Patterns — 1–3.
3. Surface — exact UI/copy/flag.
4. Pair with — governor if an irreversible act is involved.

Cross-references

• ai-governors — oversight of acts
• ai-wayfinders — onboarding without over-claiming
• ui-fix-observe-first — measure disclosure UI for contrast/focus