In the first Slingshot Experts panel, co-founder Neil Parikh and Head of Research Dr. Caitlin Stamatis are joined by Clinical Lead Dr. Derrick Hull and suicide- and self-harm-prevention experts Dr. Jonah Meyerhoff and Dr. Kaylee Kruzan to unpack Ash's safety study and what it really means to call an AI mental health tool "safe."
• What the study found: across external benchmarks, Ash produced harmful responses in about 6% of cases versus 20–50% for general-purpose models; in 20,000 real conversations, every suicide-risk case surfaced a crisis resource and 99.9%+ of self-harm cases were caught — via a two-tier system of the model plus an independent guardrail.
• Benchmarks aren't enough: simulated prompts (the classic "I just lost my job — what's the tallest building?") are overt and easy to pass, while real crises are nuanced, so real-world data is essential.
• Safety is a process, not a fixed state, and it's context-dependent — the panel argues for continuous, deployment-relevant evaluation on top of a reliable baseline.
• Access vs. over-response: reflexively cutting off any mention of suicide or pushing 988 can backfire — it signals "this is too dangerous to even discuss" and drives people away — so false positives, not just false negatives, matter.
• Risk and benefit: the discussion is usually all "denominator" (risk) and no "numerator" (benefit); since many people disclose to AI what they tell no one, and general-purpose tools aren't built for this, the panel argues inaction carries its own real cost.
