HAIER: AI as a Conversational First Responder

A mixed-methods study exploring how voice-based AI can guide untrained users through life-saving emergency procedures like choking and bleeding response.

📌PROJECT SCOPE

Client: Safety, Human Factors, and Resilience Engineering (SHARE) Lab
Timeframe: 6 weeks (part-time)
My Role: Mixed Methods Researcher, Transcriber, Data Analyst
Team: Changwon Son (PI), Seungcheol Lee (Lab Manager), Ben Alfonsin, Yedo Ji
Methods: Experiment, Usability Testing, Survey, Linguistic Analysis
Tools: SPSS, Qualtrics, LIWC-22, Excel, Premiere Pro

🪟 Project Overview

🚀 CLIENT KICKOFF
We partnered with trained first responders to investigate how well voice-based AI could provide first-aid instruction in real time. This required balancing safety, clarity, and trust in high-stress moments.

🔎 OBJECTIVES

Compare the effectiveness of ChatGPT vs. human responders in guiding first aid.
Evaluate how quickly participants can:
• Apply the Heimlich maneuver for choking.
• Use a tourniquet to stop bleeding.
Assess task performance and completion time.
Measure trust in the AI teammate and self-belief using validated survey instruments.
Understand how AI guidance affects user confidence, anxiety, and workload.

✏️ NOTES
The study involved real medical equipment and training from health professionals. The two task types — Heimlich maneuver and tourniquet application — were performed under guidance from either a human or ChatGPT.
The setting simulated real-world urgency and uncertainty.

👩🏻‍🔬 Methodology

Design: 2x2 Mixed Factorial Experiment

Participants: 56 students across AI and human coaching conditions

Tasks: Choking response, bleeding control

Measurements: Task completion time, expert video reviews, trust and workload surveys, linguistic patterns

Analysis: LIWC-22 linguistic analysis, statistical testing (SPSS)

Equipment used in the two task conditions

🤿 Deep Dive

1️⃣ EXPERIMENT SETUP

Participants worked in teams of two to complete two emergency scenarios: choking and bleeding.
Half were guided by ChatGPT (human-AI teams) and half by a human “first responder” (human-human teams) over a computer.

2️⃣ TIMING AND PERFORMANCE

We tracked how quickly participants completed the tasks and reviewed their performance with medical faculty to quantify task success. Human-AI teaming participants took less time and made fewer errors to succeed in performing the two medical tasks compared to human-human teams.

3️⃣ PERCEIVED WORKLOAD

We measured participant’s feelings of trust and workload using NASA TLX. Human-AI teaming participants reported lower workload under AI guidance in choking situations.

4️⃣ TRUST AND SELF-EFFICACY

We used Tenhundfeld’s trust scale and a self-efficacy scale. Human-human teams trusted their teammates more, but there was no notable difference in self belief between team type.

5️⃣ DIFFERENCE IN VERBIAGE

I transcribed every session and we analyzed them with LIWC-22 linguistic analysis software. Human-AI teams used more complex verbiage when communicating and communicated more than the human-human teams did, suggesting more in-depth communication with AI teammates compared to human teammates.

💭 Reflection

Industry Takeaways

Emergency-response UX must emphasize clarity, redundancy, and speed.
Voice-based AI should be paired with visual feedback for high-risk scenarios.
AI-human teams require different design scaffolding than human-human teams — especially during transition moments.

Personal Takeaways

This project taught me to look beyond outcomes and study how usability breakdowns occur in real-time.
I gained hands-on experience with transcription, linguistic analysis, and teamwork theory in high-pressure UX environments.
Designing to reduce fear, urgency, and uncertainty helped me appreciate the human side of AI-assisted interaction.