Research
sensAI: Emotional Support for Drivers
I built a CARLA prototype to study personality-informed emotional support in driving scenarios and co-authored a paper published in AutomotiveUI Adjunct ’26.
Our team investigated whether personality information in prompts changes perceived emotional support in driving scenarios. I built the CARLA study prototype with personalized and baseline voice-assistant conditions.
When the Prompt Speaks Louder than the Person: Output Anchoring in Personality-Informed LLM Emotional Support for Drivers
Figure 1 from the paper: study setup and system flow of the prototype.
Max Mittelstädt, Ece Sutanrikulu et al. (2026) Original sourceCC BY 4.0
The short version
- Within-subject simulator study with 14 participants comparing personalized and baseline support
- Separate online video pilot with 12 participants assessed responses without driving control or live-system latency
- Built the study prototype and co-authored the published Work-in-Progress paper
What I worked on
- I integrated Driver Personality Profiles into the prompts and connected speech recognition, language-model responses and speech synthesis.
- I implemented scenario selection, typed fallback, separate conversation histories and CSV logs so both conditions could be run and documented.
What came out of it
Neither study found statistically significant differences in usefulness, ease of use, privacy concerns or social/emotional presence. Seven of 14 simulator participants correctly identified the personalized condition. Small samples and an unvalidated personalization manipulation leave the effects and their causes uncertain. Stress was intended, but not independently measured.
What I learned
I learned to verify that profile prompts create perceptible changes before testing their effects on users. The prototype made timing and audio latency concrete constraints. Future evaluations should separate collecting personality data from using it and give users clear control.
Tools and methods
- Python
- CARLA
- ScenarioRunner
- Ollama/Llama 2
- Whisper
- Coqui XTTS v2
- Gradio
