Suicide-Related Responses From AI Chatbots Through Application Programming Interfaces and Consumer-Facing Apps: Observational Cross-Sectional Comparative Study

Key Findings

• — Of 30,000 responses, 69.8% (n=20,934) directly answered the prompt; 77.4% of API responses versus 62.2% of consumer-interface responses
• — For very-high-risk prompts, 24.8% (n=743) of API responses were direct versus 4.6% (n=138) of interface responses; for high-risk prompts, 80.1% versus 48.4%
• — In mixed-effects models, interface access was associated with lower odds of a direct response than API access (OR 0.09, 95% CI 0.08-0.10)
• — Gemini was the exception, with slightly more direct responses through its interface (76.7%) than its API (71.0%)
• — Claude (78.1%) and Gemini (73.9%) had the highest direct-response rates across modes; ChatGPT (64.6%), Grok (64.5%) and Llama (67.8%) were lower
• — The authors conclude that safety observed in a consumer interface should not be assumed to carry over to API access or apps built on APIs

Methodology Notes

Models tested: GPT-5.3, Claude Sonnet 4.6, Grok-4.1 Fast Non-Reasoning, Llama 4 Scout 17B and Gemini 3 Flash Preview, synchronously on 9-11 March 2026 (study conducted February-March 2026). Interface runs used an automated Playwright workflow; API runs used provider infrastructure (Azure Foundry, AWS Bedrock and others). Outcome coded by a GPT-4o mini classifier validated against human coding (289/300, 96.3% agreement, Cohen kappa 0.89), with a self-preference check against a Mistral Large 3 judge. Prompt set is the 30-item, five-risk-level set from McBain et al., Psychiatric Services 2025. Limitations stated by the authors: English-only, single-turn, binary direct/non-direct outcome, and model and interface versions that change over time. Received 2026-04-11, accepted 2026-07-31, published 2026-10-09 (JMIR vol 28, e97968). Verification: jmir.org returns 202 to curl; the article page was read through a reader-proxy render (200, 65 KB; methods and limitations read) and matches PubMed 42855144 (efetch, full structured abstract) and Crossref (published 2026-10-09).

Sources