Empathic AI in Response to Stressful Driving Situations

Introduction

Empathic AI in Response to Stressful Driving Situations is a research project from the 6th semester (SS 2025) at the Technische Hochschule Ingolstadt by Hamza Dursun, Selin Durmus, and Mona Jeske.

The project investigates whether an empathic voice assistant can reduce drivers' perceived stress, provide emotional support, and improve the acceptance of such systems after stressful driving situations.

While modern Advanced Driver Assistance Systems (ADAS) increasingly monitor the driver's state, the step beyond pure detection – the emotional support of the driver – remains largely unexplored.

Research Poster Overview

Problem Statement

Stressful driving events such as near-accidents or high cognitive load can impair both driving safety and emotional well-being. While physiological stress detection has already been researched, there is little empirical work on how voice-based AI assistants can help regulate stress and provide emotional support after such events.

This study bridges the gap between Affective Computing and In-Vehicle Interaction Design.

Main Hypotheses:

H1: Participants report lower perceived stress with an empathic AI assistant
H2: Participants perceive higher emotional support from the system
H3: Participants show greater willingness to use the empathic AI assistant in real driving situations
H4: Participants feel a higher sense of safety during stressful driving situations with AI support

Methodology

Study Design

The study used a Within-Subject Design with 20 participants (age: 25–45 years). Each participant went through a driving session with two stress-inducing scenarios in a high-fidelity hexapod simulator:

Scenario 1: Challenging Maneuver (Sharp Cornering)

A demanding curve sequence that required advanced driving skills and induced technical stress.

Scenario 2: Sudden Hazard

A critical event where a child unexpectedly appeared on the road, requiring immediate braking. The vehicle's automatic braking system was deactivated during this event.

In the Hexapod Simulator

Each scenario was conducted once with and once without the empathic AI voice assistant, with the order balanced across participants.

Voice-AI Intervention

The empathic voice assistant delivered a standardized message, about 30 seconds long, immediately after the stressful driving event. The message aimed to support emotion regulation and the perceived sense of control through calming, reinforcing, and affirming statements.

Example Message:

"That was a close moment, but you handled it very well. You stayed calm and made the right decisions."

The design is based on established psychological theories:

Emotion Regulation Theory emphasizes the importance of cognitive reappraisal and social support in stress management
The calming and slow speech output was based on relaxation techniques to reduce physiological arousal

Technical Setup

Experimental Setup

The experimental setup consisted of:

Simulation Environment: Driving route developed with IPG CarMaker and imported into the main simulation software
User Interface: Integration of visual animation and voice assistant, designed in Adobe XD, displayed on a tablet in the hexapod
Remote Control: Tablet connected to experimenter computer via TeamViewer for precise control
Communication: Smartphone in the simulator connected to lab computer via Discord

Data Collection

Physiological Data (Heart Rate)

Continuous heart rate measurement with Fitbit wristband during the entire experiment. The data was collected and synchronized via a self-developed Fitbit application.

Analyzed Time Periods:

Baseline: 2-minute period before scenario start
Stress Peak: 30-second interval immediately after stress event
Recovery: 60-second period after AI intervention or after event end

Calculated Metrics:

Stress Induction (ΔHR_stress): Change in heart rate from Baseline to Stress Peak
Recovery Amount (ΔHR_recovery): Decrease in heart rate from Stress Peak to Recovery Phase

Fitbit Heart Rate Monitoring Application

Subjective Data

After each scenario, participants filled out standardized questionnaires (7-point Likert scales):

Perceived Stress
Perceived Emotional Support
Perceived Distraction
Perceived Safety and Control

After the second scenario, a system evaluation was conducted to assess:

User Acceptance
Perceived Empathy
Trust in the Voice Assistant

Results

Physiological Results (Heart Rate)

Stress Induction Comparison

Heart Rate Change after Scenario

Both scenarios induced similar stress levels:

Curve Scenario: +4.76 bpm (SD = 5.33)
Sudden Hazard: +5.63 bpm (SD = 6.49)
Statistics: t(19) = -0.504, p = 0.6198 (not significant)

This was an intended result of the experimental design to ensure that both scenarios generated comparable physiological stress levels.

Recovery Phase (H1)

Heart Rate Recovery with and without AI

With AI: -3.20 bpm (SD = 4.42)
Without AI: -4.12 bpm (SD = 4.69)
Statistics: t(19) = -0.830, p = 0.7915 (not significant)

Result: The presence of the AI assistant did not lead to a statistically significant improvement in physiological recovery. H1 was not supported.

Subjective Results

Comparison of Questionnaire Results

H2: Emotional Support ✓ SUPPORTED

With AI: M = 5.11 (SD = 0.76)
Without AI: M = 2.42 (SD = 0.98)
Statistics: p < .001, d_z = 1.38 (very large effect)

Perceived emotional support was significantly higher in the AI condition.

System Helpfulness

With AI: M = 4.05 (SD = 0.82)
Without AI: M = 2.16 (SD = 1.05)
Statistics: p = .002, d_z = 0.87

H4: Sense of Safety (Weak Trend)

With AI: M = 4.53 (SD = 0.91)
Without AI: M = 4.00 (SD = 0.89)
Statistics: p = .403, d_z = 0.26 (not significant, trend only)

H3: User Acceptance ✓ SUPPORTED

Distribution of User Preferences

85% (17/20) of participants reported feeling better supported with the voice assistant. Only 15% (3/20) saw no difference.

Qualitative Feedback

What Participants Appreciated:

Positive Qualitative Themes

Voice Tone & Empathy:

"Felt genuinely understood." (12/20)
"The AI sounded calm and reassuring, which reduced my anxiety." (9/20)

Timing & Contextual Fit:

"Receiving the message immediately after the event felt just right." (15/20)

What Was Criticized:

Negative Qualitative Themes

Conciseness & Clarity:

"The message was too long and repetitive." (5/20)
"Shorter prompts with bullet points would be ideal." (7/20)

Timing Issues:

"Sometimes it interrupted my focus—would help if I could delay." (4/20)

Customization & Control:

"Would be better if it addressed me by name or adapted to my driving style." (8/20)
"An option to switch to visual cues would improve usability." (6/20)

Discussion & Interpretation

The study shows a dissociation between physiological measurements and subjective evaluations:

Physiological: No significant improvement in heart rate recovery
Subjective: Strong and significant improvements in emotional support and system acceptance

Possible Explanations:

Cognitive Reappraisal: The empathic AI assistant likely enabled a reinterpretation of the stressful event, reducing the perceived threat even if physiological arousal remained unchanged.
Expectation and Novelty: Interaction with an empathic voice agent might have triggered positive affect through expectation violation and novelty.
Attention Allocation: The AI's verbal support might have redirected participants' attention from internal stress signals to external reassurance.
Self-Report Sensitivity: Subjective measures (Likert scales) might be more sensitive to perceived social presence than physiological signals like heart rate.

Limitations & Future Work

Identified Limitations:

Physiological Measurement: Only Heart Rate (HR) was measured. Heart Rate Variability (HRV) would provide more robust insights into autonomic nervous system regulation.
Experimental Design: No full balancing of scenario type and intervention. Future studies should use Latin Square or fully balanced designs.
Message Length & Timing: Several participants found the message too long or repetitive. Delayed empathic reactions could reactivate stress instead of reducing it.
Risk Compensation Effect: Reassuring feedback could lead to overestimation of one's abilities or reduced vigilance.
Sample Size: N=20 is consistent with exploratory HCI studies but should be confirmed with larger, more diverse samples.

Future Research Directions:

Sophisticated Sensors: Integration of HRV, Electrodermal Activity (EDA), and other indicators
Adaptive Systems: Dynamic adaptation of content, timing, and modality (voice, haptic feedback, ambient light) based on real-time affect sensing
Longitudinal Studies: Observation over weeks/months to assess long-term effects on trust and well-being
Personalization: Consideration of individual personality traits and preferences

Design Implications

Based on the results, the following design recommendations arise for empathic in-vehicle systems:

User Control & Customization: Adjustable message length, volume, and feedback modality (visual vs. auditory)
Multimodal Integration: Combination of voice feedback with subtle visual or haptic cues (e.g., calming animations, ambient lighting)
Context-Aware Safety Protocols: Context awareness to avoid false reassurance in actually dangerous situations
Conciseness & Timing: Shorter, better-timed, and adaptive support

Ethical Considerations

While empathic AI can improve perceived emotional support, it typically provides standardized responses that only simulate empathy. This raises ethical questions regarding:

Authenticity & Transparency
Potential for Manipulation
Parasocial Bonds
Overestimation of Own Abilities through permanent AI support

These aspects require careful design and responsibility.

Conclusion

The empathic, voice-based AI assistant did not lead to significant improvements in physiological recovery but was highly appreciated by participants on a subjective level.

Key Findings:

✓ Significantly higher emotional support (H2 supported)
✓ High User Acceptance (H3 supported, 85% agreement)
✗ No physiological improvement (H1 not supported)
~ Slight trend in sense of safety (H4 weak trend)

The results provide actionable insights for the development of emotionally responsive in-vehicle AI systems and emphasize the importance of conciseness, adaptivity, user control, and multimodal design.

Future systems should provide shorter, better-timed, and adaptive support, possibly integrated with real-time affect sensing.

Technical Details & Repository

The project included the development of a custom Fitbit application for precise heart rate measurement and synchronization with experimental events.

GitHub Repository: https://github.com/hd2386/fitbit

The application enabled:

Continuous HR data collection
Manual logging of start times (hour, minute, second) of each experimental phase
Precise temporal alignment of event markers with physiological data
Data retrieval via the Fitbit API

Team & Acknowledgments

Project Team:

Hamza Dursun (had2386@thi.de)
Selin Durmus (sed0245@thi.de)
Mona Jeske (moj9886@thi.de)

Special Thanks:

Claus Pfeilschifter for simulator support
Markus Weißenberger for technical support at CARISSMA
Prof. Dr. Ignacio Alvarez (Technische Hochschule Ingolstadt) for valuable feedback

Institution: Technische Hochschule Ingolstadt (THI) Esplanade 10, 85049 Ingolstadt, Germany

This project was conducted as part of the course FWS/SDUT 2025 at the Technische Hochschule Ingolstadt.

The complete bibliography and detailed method description can be found in the full Research Paper.

Keywords: Empathic voice assistant, driver stress, in-vehicle interaction, affective computing, user study, human–AI interaction, physiological sensing, heart rate