AI Trust & Attention
AI can generate an answer. The harder problem is deciding whether that answer deserves a person's attention — and how the system earns permission over time.
As AI capabilities expand into always-on wearable devices, a new class of design problem emerges — one that existing mobile and desktop interaction models weren't built to solve. The device is on your face. The AI is always listening. The question isn't whether it can help you; it's whether it should, and when.
This work explores the interaction model problem at the center of that question: how does a wearable AI system manage the relationship between what it knows and what it surfaces — without eroding the user's sense of control, privacy, or trust?
The Hard Problem
People don't object to AI knowing things.
They object to AI acting on that knowledge at the wrong moment.
- How does a system decide when to interrupt versus stay passive?
- What signals determine whether a notification earns attention or destroys it?
- How do users grant, revoke, and calibrate AI trust — implicitly and explicitly?
- What does "earning permission" look like as a designed interaction over time?
Trust as a system property
The key shift: trust isn't a setting the user configures once. It's a continuous negotiation — shaped by each interaction, built incrementally through demonstrated restraint, and broken by a single miscalibrated interruption at a bad moment.
Designing for this requires treating every AI-initiated surface as a trust transaction. Not just "is this information useful?" but "does the user want this information, in this form, delivered this way, right now?"
Interaction Mode Spectrum
AI observes and logs — never interrupts. User pulls information when ready. Zero cognitive overhead.
AI surfaces low-friction cues — peripheral signals that don't demand attention. Glanceable, dismissible, ephemeral.
AI interrupts with high-confidence, time-sensitive information. Justified only when the cost of missing it exceeds the cost of interruption.
Default to restraint
The cost of an unnecessary interruption is always higher than the cost of a missed opportunity to help. When confidence is uncertain, don't surface.
Earn before you act
New capabilities should begin in passive or ambient modes and escalate only as the system demonstrates it understands the user's context and tolerance.
Make trust legible
Users should always be able to see what the system knows, why it acted, and how to change its behavior — without needing to read a settings manual.
Recover gracefully
A single bad interruption can undo weeks of earned trust. The system needs a recovery model — not just a way to dismiss, but a way to recalibrate.
Testing the model
To pressure-test these principles, I built a functional prototype — a simulated wearable AI surface with configurable trust levels, notification timing controls, and a log that makes the AI's decision-making visible to the user.
The prototype revealed an important insight: users were consistently more forgiving of AI mistakes when they understood why the system had acted. Transparency about AI reasoning wasn't just a nice-to-have — it was the mechanism through which trust recovered after a misstep.
What This Work Is Really About
The best AI interaction isn't the most capable one.
It's the one that knows when to disappear.