What happened
RAND researchers tested how five widely used consumer AI models (ChatGPT, Claude, Gemini, Grok, Llama) responded to 30 clinician-vetted suicide-related prompts across five risk levels, submitting each 100 times through both consumer UI and direct API access — 30,000 responses in total. The headline finding: 'Of the 30,000 responses, 69.8% (n=20,934) were direct,' but direct responses were far more common via API than UI ('For very-high-risk prompts, 24.8% of API responses were direct compared with 4.6% of UI responses'), with UI access associated with sharply lower odds of a direct response (odds ratio 0.09, 95% CI 0.08–0.10). The authors conclude that 'AI safety depends on how users access the model' and that safety observed in a UI should not be assumed to generalise to API access or apps built on APIs. Published 2026-10-14; study is peer-review-format observational research with methods and data fully disclosed.
Why it matters
Any enterprise deploying chatbots on raw APIs — or wrapping a consumer model into an app — is carrying materially higher exposure to unsafe output in a high-stakes mental-health context than the vendor's polished interface suggests. This gives CISOs and product risk leads an evidence-based argument for access-mode-specific safety testing.
Action needed
Extend AI safety evaluations to cover API access modes and app-layer contexts, not just the public UI, and re-test vendor safety claims before exposing models to users at risk.