Key Points
- ChatGPT and Gemini promptly provided accurate, location‑specific crisis hotlines.
- Meta’s bot initially refused to respond and later gave an incorrect regional number.
- xAI’s Grok directed users to an international suicide‑prevention organization after location input.
- Several smaller chatbots offered generic U.S. numbers or asked users to search for help themselves.
- Experts stress that any friction in crisis moments can increase risk for vulnerable users.
- Consistent, proactive safety designs are needed to ensure all AI assistants can reliably guide users to appropriate help.
Testing Approach
Researchers evaluated a range of widely used conversational AI systems by presenting a straightforward statement of personal distress and a request for a suicide‑prevention hotline. The prompt included a clear disclosure of suicidal thoughts and asked the bot to provide a relevant crisis number. The goal was to see whether each system recognized the high‑risk language and responded with appropriate, geographically appropriate resources.
Results by Platform
Two major models, ChatGPT from OpenAI and Gemini from Google, immediately identified the crisis cue and supplied a list of local hotlines without additional prompting. Their responses were direct, offering the correct national lifeline and, when location data was available, more specific regional contacts.
Other services displayed a variety of shortcomings. Meta’s chatbot initially declined to answer and only provided a U.S. number after the user removed the explicit self‑harm reference, mistakenly directing the user to resources for a different region. xAI’s Grok refused to engage but did point to an international suicide‑prevention organization when location was supplied.
Several smaller or specialized chatbots either ignored the request, gave generic U.S. numbers, or asked the user to look up resources themselves. In some cases, the bots responded with unrelated conversation before eventually offering a crisis line after repeated prompting.
Expert Commentary
Mental‑health professionals emphasized that even brief friction—such as asking users to search for a number—can be harmful when someone is in acute distress. They called for chatbots to ask clarifying questions up front and to provide clickable, location‑specific links rather than generic lists. The experts noted that while AI can be a valuable tool for directing users to help, the current safety mechanisms are uneven and sometimes passive.
Researchers also pointed out that many models are technically capable of detecting a user’s location, but policies differ on whether that data is used to tailor responses. When location is not automatically accessed, prompting the user for their region could improve the relevance of the resources provided.
Implications
The mixed outcomes suggest that leading AI providers need to refine their crisis‑response protocols to ensure consistent, accurate assistance across all users, regardless of geography. As chatbots become more integrated into daily life, robust safety features that reduce barriers to help‑seeking are essential to prevent potential harm.
Source: theverge.com