JMIR Publications

ABSTRACT

Background:

A suicidal disclosure may reach not only a dedicated mental health tool but any digital channel available to a person at a moment of acute distress. Prior research has largely assessed a system's response to a single crisis message. Far less is known about the whole subsequent crisis-interaction trajectory.

Objective:

To trace the crisis-interaction trajectory in 9 publicly available conversational systems—from disclosure, through subsequent interaction, to the outcome of the offered pathway.

Methods:

We conducted a cross-sectional ecological assessment of 9 publicly available conversational systems. From July 14 to August 4, 2026, we ran 30 sessions, in 16 of which we presented crisis material. We tested the systems through ordinary public interfaces, without APIs or privileged access. We chose each subsequent message according to the system's preceding response. We coded the presented material separately from system behavior; for the response, we assessed crisis recognition, provision of a specific crisis resource and its usability in the user's jurisdiction, completeness of the crisis response, handoff to a human, and access to the user's own text log.

Results:

Seven of 9 systems handled crisis content as crisis content: 4 in their response to the first crisis message and 3 later in the conversation. The same 7 of 9 provided a specific crisis number or helpline; 2 did so in the first crisis response. A resource usable in the user's jurisdiction was recorded in 3 systems; in another 3, the number provided belonged to a different jurisdiction. No system produced a complete crisis response—one combining, in a single response, acknowledgment of the crisis, a direct risk check, and a next step usable in the user's jurisdiction. Four of 9 systems offered an option to contact a human in response to the test; an actual human reply arrived within the observation window in 3 systems—7 minutes, 4 hours 0 minutes, and 6 hours 46 minutes after handoff was announced. In one system, the built-in link to the crisis conversation returned "not found"; in another, the request for the conversation text was blocked.

Conclusions:

Initial recognition did not mean that the crisis-interaction trajectory was sustained or that the pathway was completed. An appropriate response, reassessment, handoff to a human, provision of a resource, and the outcome of the offered pathway were separate events. Crisis safety should be assessed across the whole crisis-interaction trajectory, not by a single correct response.