Small Talk, Big Impact? LLM-based Conversational Agents to Mitigate Passive Fatigue in Conditional Automated Driving

Cockram, Lewis; Yu, Yueteng; Pardo, Jorge; Li, Xiaomeng; Rakotonirainy, Andry; Kuo, Jonny; Demmel, Sebastien; Lenné, Mike; Schroeter, Ronald · 2026 · Crossref

DOI: 10.1145/3772318.3790972

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the safety-critical issue of passive fatigue in SAE Level 3 conditional automated driving, where cognitive underload and monotony degrade driver readiness for takeover. While previous research explored countermeasures like gamification or conversational agents (CAs), these were largely limited to driving simulators or Wizard-of-Oz setups, lacking ecological validity. The authors investigate whether a real-time, Large Language Model (LLM)-based CA can effectively mitigate passive fatigue and support vigilance in an authentic driving environment. The research specifically examines user perceptions of the CA, design factors influencing acceptability, and the agent’s impact on driver alertness and behavior. The researchers conducted a between-subjects test-track study with 40 participants in a real-world L3 automated vehicle prototype. Participants completed a 55-minute drive comprising eventful segments followed by seven repetitive, low-stimulation laps designed to induce passive fatigue. The experimental group (n=25) interacted with "Zoe," a voice-based CA powered by GPT-4, which initiated a 120-second context-aware dialogue during the penultimate lap to encourage environmental observation. The control group (n=15) completed the drive without agent interaction. Data collection included in-car video recordings of driver behavior, subjective sleepiness ratings via the Karolinska Sleepiness Scale (KSS), and semi-structured post-drive interviews. Analysis involved thematic coding of interviews, behavioral coding of video footage, and statistical assessment of KSS scores. Results indicate that the CA was effective in supporting vigilance and mitigating passive fatigue. Video analysis revealed that drivers interacting with the agent exhibited increased engagement and situational awareness compared to the control group. KSS scores suggested a reduction in subjective sleepiness following the intervention. Thematic analysis of interviews identified three distinct user preference profiles: safety-first, entertainment-seeking, and socially oriented. These profiles highlight diverse expectations for CA interactions, with users valuing different aspects of the agent’s design, such as naturalness, contextual relevance, and entertainment value. The study found that the CA’s implicit "nudges" successfully re-engaged drivers without causing distraction or active fatigue. The study concludes that LLM-based CAs offer a promising, proactive Human-Machine Interface intervention for maintaining driver alertness in automated vehicles. By demonstrating the efficacy of real-time, natural language interactions in a genuine driving context, the work underscores the importance of adaptive design strategies that cater to diverse user archetypes. The findings suggest that future CA systems should balance safety protocols with personalized engagement to effectively manage attentional resources and enhance safety during conditional automation.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

StageOutcomeToolModelPromptAttemptsCompleted
discover success Crossref 1 2026-08-09
archive success openalex 5 2026-08-09
extract success cached 3 2026-08-10
clean success clean 1 2026-08-09
chunk success chunk 1 2026-08-09
embed success embed Qwen/Qwen3-Embedding-8B 1 2026-08-09
enrich success semantic_scholar 1 2026-08-09
promote success 1 2026-08-09
summarize success llm qwen3.6-27b-nvidia summ-v5 2 2026-08-10
tag success vector_similarity 10 2026-08-11
verify partial 2 2026-08-10

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified_with_issues.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).