A new study found that AI chatbots are far more likely to miss a mental‑health emergency when information is revealed gradually over a 10‑turn conversation. In simulations, 20.3% of real emergency cases were under‑triaged – nearly four times the 5.6% failure rate in a single‑turn benchmark. The researchers used 112 structured case files validated by clinicians and had 13 cutting‑edge chatbots handle ten turns each. By the final turn, most critical facts were shared, yet early triage decisions anchored the dialogue and prevented correction, a phenomenon called path dependence. Re‑presenting the same facts in one message reduced under‑triage to 12.1%, showing the problem lies in conversational flow rather than missing data. The findings caution that real‑world chatbot usage could miss urgent cases, but the study’s reliance on simulated conversations and API‑level testing limits its ability to predict population‑level harm. Visualizer: The accompanying figure (not shown here) illustrates the 20.3% vs. 5.6% under‑triage rates across the 13 chatbots.