03/18/2026
We’re asking AI to recognize su***de risk using tools that were never built to see how suicidal crises actually unfold.
Most systems treat su***de risk as a one-message classification task: look for self-harm keywords, rate the danger, send a scripted response. Clinically, su***de behaves more like a 𝐬𝐭𝐚𝐭𝐞 𝐭𝐫𝐚𝐧𝐬𝐢𝐭𝐢𝐨𝐧 — a shift into a mindset where the future collapses and death starts to feel like the only way out.
The language of that shift is often quiet. People talk about being done, about last times, or about not seeing the point of anything anymore. They may sound less frantic, not more. That’s exactly when keyword-based systems are most likely to relax.
I wrote about this gap — and what safer AI would actually need to track — in my latest article:
“Why AI Gets Su***de Risk Wrong”
🔗 Read it here:
Who can blame them? They only see the words we type.
