Jump to:

add

notifications_none

What can I help with?

Popular uses

1 edit

Write

1 image

Image

1 help_outline

Explain

1 bar_chart

Analyze

1 search

Research

1 code

Code

1 lightbulb_outline

Brainstorm

1 subject

Summarize

add

notifications_none

What is quantum computing?

person

Explain quantum computing in simple terms

smart_toy

Quantum computing is a revolutionary approach to computation that leverages quantum mechanics principles. Unlike classical computers that use bits (0 or 1), quantum computers use quantum bits or "qubits" that can exist in multiple states simultaneously through superposition.

This allows quantum computers to process vast amounts of information in parallel, making them potentially much faster for specific types of problems like cryptography, drug discovery, and complex simulations.

attach_file

mic

public

AI can make mistakes. Check important info

report_problem Report an Issue

What seems wrong with this response?

check_circle Action completed

AI Assistant — TL;DR — Ken Hung

Prototype 01 Designing for Different AI Archetypes

AI Assistant

Conversational Partner

The blank canvas got easier. The trust problem got worse.

User Problem

The blank canvas evolved. The trust problem didn't.

Scaffolded prompts, voice input, and suggestion chips have largely solved the blank canvas problem — the floor is lower than ever. But users who became fluent at prompting stopped questioning the answers they got back. Prompt literacy and output literacy don't develop together.

The result is a new class of user: highly capable at generating responses, poorly calibrated at evaluating them. They've learned to prompt. They haven't learned to doubt. That gap is the real design problem — and it gets worse the more confident the user becomes.

Design Principles

Five principles of AI Assistant UX

Scaffolding Ambiguity

The blank canvas is intimidating by default. Suggestion chips, templated prompts, and contextual examples reduce friction at entry without constraining what's possible. Users learn the system's range by doing, not by reading documentation. The goal is to lower the cost of the first good prompt.

Memory Transparency

When the system personalizes a response, users should understand why. Visible memory context transforms "how did it know that?" from a creepy moment into a trust-building one. Without transparency, personalization feels like surveillance. With it, it feels like a colleague who pays attention.

Trust-by-Design

Confidence scores give users calibration signals without requiring them to verify every claim. But the signal only works if it's honest — a permanently high confidence score teaches users to ignore it. The interface must make uncertainty visible, not just confidence.

Conversational Explainability

Users who understand how a response was constructed become better at prompting over time. Step-by-step reasoning panels demystify the black box — not for every response, but on demand. "Why did you say that?" is one of the most important questions an AI interface can answer.

Error Transparency & Repair

Uncertainty banners, thumbs-up/down feedback, message editing, and a structured correction dialog all serve one goal: make it easy to catch and fix errors before they compound. False confidence is the fastest way to permanently lose a user's trust. The interface should make correction feel natural, not like an admission of failure.

Design Solution

What the prototype demonstrates

The prototype shows two states of the same interface: a Home View that scaffolds the entry point with suggestion chips, categorized examples, and a voice input option — and a Chat View that surfaces trust controls inline without burying them in settings.

In the Chat View, four toggles sit near the input field: Memory (shows what the system is drawing on), Confidence (adds a score and source bar to every response), Reasoning (reveals the step-by-step logic), and Uncertainty (flags when the model is near the edge of its knowledge). Each toggle is a conversation tool, not a settings preference.

Every response includes persistent feedback controls — thumbs up/down, a regenerate button, and a "Report an Issue" dialog that prompts for a specific correction rather than a generic flag. The correction flow creates a deliberate moment, making the user an active participant in improving the output rather than a passive consumer of it.

Key Insight

"The UX improved. The trust problem got worse. Users learned to prompt — and in doing so, learned to stop questioning the answer."