Which Aspects of a Conversation Keep Drivers Awake? An Exploratory Analysis of Driver–AI Dialogue

Driver drowsiness during prolonged, monotonous driving remains a major road safety risk, and conventional audio-haptic alerts often fail to counter it. This exploratory study examined whether specific features of driver-AI conversation are associated with reduced drowsiness. Ten participants drove a 50-minute high-fidelity simulated highway session, with a five-minute unscripted voice-based conversational intervention with ChatGPT-4o introduced after 30 min. Sleepiness (KSS) was measured before, immediately after, and at three post-intervention timepoints. Conversational data were analyzed using lexical, sentiment, and embedding-based semantic measures, then correlated with change in KSS. Mean KSS decreased significantly following the intervention and remained below baseline for 15 min. Compound sentiment, fear-induced emotional arousal, and AI-to-human novelty ratio showed the strongest, directionally stable associations with reduced drowsiness, while coherence, drift, and topic breadth showed weaker associations. Given the small sample and correlational design, findings should be interpreted as preliminary, informing future controlled studies on designing engaging in-vehicle conversational agents.