Designing AI that knows when not to act

Context

Motorcycling is a high-risk, attention-saturated activity.

Riders continuously process:

• speed
• road conditions
• navigation
• risk

Most rider technologies fail by:

• increasing visual load
• interrupting at the wrong time

After experiencing a motorcycle accident, I reframed the problem:

Not how to add information,
but how to intervene without adding cognitive load.

The problem

This was not an interface problem.

It was human–AI coordination under constraint:

• attention already saturated
• reaction time critical
• interruptions risky

The core challenge:

How does a system help without competing for attention?

The shift

I reframed the system from information delivery to intervention control.

From:

• more signals → fewer, higher-value signals
• constant assistance → selective intervention
• visibility → timing

In safety-critical systems:

Value comes from when the system acts, not how much it shows.

The system

Gia operates as a real-time intervention system:

• Observes continuously
• Predicts high-risk moments
• Intervenes selectively
• Stays silent when confidence is low

The system prioritizes what not to surface.

Design & prototyping

I focused on identifying when riders have cognitive availability versus overload.

This included:

• mapping high-risk vs low-attention moments
• testing voice timing under real riding constraints
• evaluating when intervention improves vs disrupts

Every decision was evaluated against:

Does this reduce cognitive burden in motion?

Voice

Voice was the primary interface.

When eyes and hands are occupied, spoken cues preserve situational awareness.

• short
• contextual
• timed to moments of availability

This was not conversational AI.
It was controlled intervention.

Design decisions

• Selective intervention over constant feedback

  The system only surfaces high-value signals.

  Result: higher trust and lower cognitive load.

• Haptics and voice over visual interfaces

  Reduced reliance on visual attention.

  Result: safer interaction under motion.

• Timing over information density

  Optimized when signals appear, not how many.

  Result: improved reaction and reduced distraction.

• Behavioral adaptation over configuration

  System learns from rider behavior instead of setup.

  Result: lower friction and more natural integration.

• Explicit failure states over false confidence

  Defined how the system behaves under uncertainty.

  Result: increased reliability and trust.

Tradeoffs

Designing for safety required restraint:

• Less information vs safer interaction
• Fewer features vs higher trust
• Conservative outputs vs perceived intelligence

We prioritized reliability over capability.