Design research - AI
Correct me if I am wrong?!
TrinkTracker, an AI-powered system for monitory daily drinking practices. The TrinkTracker is capable of recognizing six types of beverages based on a photo and generating a data report summarizing people’s drinking behaviors. A two-week field study with 14 participants has been conducted to understand user perception, experience and trust on AI predictions by comparing two conditions: TrinkTracker and a baseline system.
My Role
Research, Analysis, Product Design, Interaction Design
System Design, User-Testing
Tools
HTML, CSS, image recognition,
Design Thinking
ACM LibraryDate
2023
Overview
Problem
Unexpected AI results may lead to frustration, disappointment, and technology abandonment, causing mistrust or over-trust.
Goal
To seamlessly integrate AI into our daily life, this work aims to investigate how AI outputs influence people’s perceptions and experiences in their everyday practice, particularly when they are given the opportunity to correct AI mistake.
Research
Introduction
Prior research has primarily focused on understanding how AI practitioners or designers make sense of AI’s outputs with the goal of improving AI performance and user experiences. Yet, little research has been explored on the understanding of end users’ perceptions and experiences with AI in everyday practice. This work explores whether providing opportunity for users to correct AI errors influences their perception and experience.
Prototype
TrinkTracker
TrinkTracker includes two components: (1) a data tracker detecting the type of beverage based on a photo, and (2) a data report generating a weekly summary of their drinking behaviors over time. In the system, users can review the output generated by the classifier and decide whether they want to modify the output.
Field study
A field study was conducted with 14 participants to understand user perception, experience, and trust in AI predictions by comparing two conditions: TrinkTrackeker and a baseline system where users cannot modify mistakes made by AI.
Results
TrinkTracker findings
The results showed that people exhibited higher levels of trust when they had the opportunity to correct the AI’s mistakes. More interestingly, the occurrence of AI mistakes sparked people’s curiosity, prompting them to speculate the underlying reasons for these errors. These findings provide valuable insights into understanding how users perceive AI mistakes in the context of everyday practices. Particularly, this work highlights the importance of user involvement in the AI feedback loop, as it positively impacts user trust and encourages a deeper understanding of the technology.