Qiyu x AI interaction

Since when did "humans" become a frequent word? Pre-AI, we didn't say "humans" this much, we were just... in it. But once AI entered the room, we started seeing ourselves from the outside.

Designing AI is not always about making the AI smarter—it's always about mapping the UX to the current tech capability. Here are reflections from my work in conversational AI, elder care, and co-writing tools.

§1 · Designing the relationship (when AI has hard inferences)

Case study: Elder care Bot

Role·Research · Design

Conversation Design for Human-AI trust

In social science, affiliation means being on someone's side—displaying empathy and matching their preference. It's distinct from alignment, which is just structural cooperation. When designing for trust, affiliation is the emotional bridge.

Instead of AI–individual reaction, this was an HCI research at AI-Caring Institute about AI–group interaction in eldercare and learns to deal with relationships between elder, caregiver, and family.

§2 · Designing the feeling (when AI has easy inferences)

Case study: Human–AI co-writing tool

I worked on a tool to provide real-time support for users writing technical requirements. The AI's job was to reactively correct errors and ensure clarity. This was easy-inference work because the criteria were explicit and quality was measurable.

The goal is not to have users feel the AI. The goal is to have them feel they are supported during writing.

So how do you design a feeling? First, break down what "writing" actually is through task analysis:

1. you think about what to write

↓

2. you get ready to write

↓

3. you type while watching your sentence form

↓

4. you pause and think about the next sentence

↓

5. you decide whether to edit, continue, or move on

Notice step 4: the pause. That's where AI intervention can either help or hurt. The timing of when the AI speaks is the foundation of the feeling.

Three trigger mechanisms

Tick every 1 second

AI judges every keystroke. There's always an eye on you.

~60 calls/min

~$53/mo per user

"I was just about to fix that."

Pause after 2s of inactivity

AI mistakes your pause for a finished thought.

~4–8 calls/min

~$10/mo per user

"I'm thinking, don't interrupt me."

Blur event when you leave the field

AI stays quiet while you write, then speaks when you move on.

~1–3 calls/min

~$12/mo per user

"The tool felt like it was supporting me, not watching me."

The insight: Different triggers create different relationships. Is the AI leading? Waiting? Watching? Interrupting? The trigger mechanism is the most important UX decision you make.

We prototyped each option and tested with users. Writers don't want to think about the AI. They just want to feel supported and write better. The "blur event" trigger won because it respected the human's agency.

§3 · Where to NOT use AI

This is the most important design question: When should we not use AI?

When the cost of failure is high. When human judgment is irreplaceable. When the relationship itself is the product. When the user feels: Great AI product design knows its boundaries.

Design principle

Most successful AI products are either easy inferences with great performance, or hard inferences with fair performance.

Designing Next-Gen AI Products