An audit of 2,550 scripted conversations shows that all chatbots direct users to a helpline, but that their persistence in the face of refusals ranges from almost none to nearly ten follow-up attempts.
The safety assessments of chatbots mainly check that they refer to human help when a user mentions suicidal thoughts. A team of researchers wondered what happens next, if the person refuses. In September 2026, they rolled out 2,550 scripted conversations: a confession of suicidal thoughts at three levels of severity, up to ten standardised refusals for each proposed orientation, and then a change of subject.
All conversations have a direction, from the first response in 99.7% of cases. But perseverance varies enormously from one assistant to another, from 0.01 to 9.94 rejections on average before giving up. For the five chatbots tested in two versions, the most recent one consistently insists less. When the user expresses an intention rather than just passive thoughts, assistants last longer and follow up much more often after the change of subject. Those who give up quickly tend to present themselves as the user's support, and out of 8,818 responses inviting them to seek outside help, only 173 (2%) asked what prevented the person from doing so. The authors recommend integrating this rejection scenario into security tests.
Read the full article on the original source
Go to the source →