Surfacing Suicidal Risk Through Simulated Social Interaction: Per-Person Language Model Agents as Communicative Stress Tests

Suicidal risk may be encoded in everyday communication patterns but diluted in routine digital interactions. We introduce a method for surfacing this latent signal: training per-person language model agents on individuals’ authored text (the on-screen text each participant typed, captured whenever a keyboard was visible in screenshots) and placing those agents in simulated social interactions—a “communicative stress test.” Using data from 79 adults with recent suicidal ideation, we fine-tuned individual LoRA adapters on Qwen3-8B using each participant’s authored text, then placed agents in standardized conversations with probe personas. Agent-generated risk language was associated with EMA-measured suicidal ideation ( r = .576, p < .001), with a single neutral small-talk probe performing nearly as well ( r = .551). A shuffle control confirmed the signal is person-specific ( r = .071 when adapters were mismatched), and automated descriptions of participants’ general smartphone activity produced no signal, confirming specificity to interpersonal communication. A prompt ablation demonstrated partial robustness to removal of disclosure-encouraging language ( r = .430). This proof-of-concept demonstrates that simulated social interaction can amplify latent vulnerability signals, bridging digital phenotyping, generative AI, and suicide theory.