Adaptive De-escalation Trainer: Piloting a RAG-Enhanced, Emotionally Modulated AI Simulator for Police Training

Effective police de-escalation training requires realistic practice but is limited by the scalability and consistency of traditional methods. We present the Adaptive De-escalation Trainer, which combines a large language model with retrieval-augmented generation for scenario grounding and text-to-speech that modulates prosody based on officer keywords. We hypothesized that the simulator would yield high conversational coherence and responsiveness (H1), emotional realism (H2), keyword-driven emotional shifts (H3), and viability (in terms of latency and workload) (H4). In a study with three officers, high average ratings for conversational flow ( M = 6.00), validity of persona ( M = 5.67), and logical responses ( M = 6.67) were found, supporting H1; high scores on emotional realism ( M = 5.75) supporting H2; positive AI reactions after “success” phrases offering support for H3; and an average turn latency of 4.3 s with moderate NASA‑TLX scores, confirming H4. These findings demonstrate feasibility and inform future work on latency reduction, scenario expansion, and advancing AI-based training.