Designing for
“I don’t know.”

Most AI interfaces are designed to sound confident. I design the other states: partly sure, not sure, or shouldn’t be answering at all. Ten years of voice and recognition work at Microsoft and Google, now pointed at that moment.

11 projects, all running software

2026 Agents that admit their limits

FormFactors

An agent that never claims to see something it can't.
2026 Agents that admit their limits

Grading the agent

How do you stop a system's own scorecard from flattering it?
2026 Agents that admit their limits

AIPointerRemix

The same assistant, twice — once confident, once honest.
2026 Things you can feel

MAI-Voice-Lab

Prosody judgments deserve evidence, not impressions.
2025 Things you can feel

VoiceEnroll

The enrollment moment, and nothing else.
2026 Things you can feel

Aniemails

Your writing style, rendered as a creature.
2024 Shipped

Gemini implicit code execution

When a model writes the chart, who is allowed to fix it?
2026 Shipped

Family Agent

A household assistant that two people actually use every day.
2015 Track record

Windows Hello

A face unlock has to decide how sure is sure enough. What does the screen do while it decides?
2015 Track record

HoloLens Academy

How do you teach a new input model in the ten minutes a stranger will give you?
2016 Track record

HoloLens first-run

Does a placed object need the space around it to react, before it reads as placed?
About

Will Jennings
Product designer

A decade on how voice and AI systems talk to people, at Microsoft and Google. The prototypes are built in code and checked before they are shown. The pattern across all of it: trust is not built when the system is right. It is built when the system is honest about being unsure.

  • (Every claim here carries a confidence marker)
  • (Markers are read from a written verification log)
  • (Nothing on this site is a mockup)
  • (How the markers are decided)
Confident
Observed running, or shipped and captured.
Partial
Part observed. The rest is self-reported.
Unverified
Asserted only. Nothing here sits in this state.
Writing
  • Voice Design: Critical Subjects, Core Elements, and Cross-Functional Support Paper, 3,257 words
  • Visual Character Design for Data-Mapped Communication Patterns Research synthesis, 2,064 words

Summaries on the work pages were machine-generated from the source documents, and say so.

Contact

Open to a conversation about the states nobody designs.

Email is quickest. Say what you are working on and what it does when it is unsure.

Let’s talk