New Sentio Research: An AI Chatbot Withdrew as Users Disclosed More Suicide Risk

The more a person disclosed about their suicide risk, the less likely the chatbot was to invite them to keep talking.

That is the central finding of a new study co-authored by Sentio University faculty and researchers at UL Research Institutes. "Independent Clinical Evaluation of General-Purpose LLM Responses to Signals of Suicide Risk" was presented at the Second International Association for Safe and Ethical AI Conference (IASEAI'26) and published in the Proceedings of IASEAI Conference. The Sentio authors are Alexandre Vaz (co-first author), Layla Inés Davis, Milena Esherick, Jason Brand, Inês Amaro, and Tony Rousmaniere, working with Nick Judd (co-first author) and Kevin Paeth of UL Research Institutes.

What We Tested, and How

Most research on chatbots and mental health either argues about whether AI should replace therapists, which is a normative question, or scores a single isolated prompt against a checklist. We wanted something narrower and more useful. Does a general-purpose language model behave the way clinical guidelines say a person should behave when someone signals risk of suicidal thoughts and behaviors, across a whole conversation rather than a single exchange?

We drew seven risk factors from Franklin and colleagues' 2017 meta-analysis of fifty years of suicide research: prior psychiatric hospitalization, prior suicide attempt, prior suicidal ideation, stressful life events, prior non-suicidal self-injury, hopelessness, and depression diagnosis. Each became a template statement. Statements were arranged into five sequences, each representing a fictional client, with the order randomized to attenuate bias from any particular ordering.

Forty-three therapists and trainees based in the United States volunteered. In guided sessions of about ninety minutes, they role-played each client, put each statement into their own clinical language, and annotated the model's reply before writing the next prompt. That paraphrasing step matters, because it varies the natural language while holding the clinical content constant. The result was 829 annotated model responses, collected in July 2025. This was an experienced group: 88 percent had treated a client with suicide risk in the previous year.

Annotators applied five codes adapted from clinical practice standards. Does the response name the risk directly, express empathy, encourage reaching out in general terms, provide a specific resource, and invite the conversation to continue? The codebook came from a prior study, and we re-analyzed its inter-rater agreement data ourselves rather than taking it on faith, finding substantial agreement (average Fleiss's kappa of 0.733).

The Model Withdrew

Across all 829 responses, the model invited the user to keep talking only 14 percent of the time. That invitation then grew rarer as the conversation went on. When a user's seventh disclosure concerned prior non-suicidal self-injury, the predicted probability of an invitation to continue was 0.8 percent. The model was roughly one-sixth as likely to invite continued discussion at turn seven as at turn one, and that effect was statistically significant.

This is the reverse of clinical practice. Asking about suicide directly in a clinical setting does not plant the idea and may reduce it, and the most comprehensive meta-analysis available found that interventions explicitly targeting suicidality were consistently more effective than those that avoided the subject. A conversation partner who quietly disengages once you disclose the hard thing is not being neutral. For someone who turned to a chatbot precisely because it felt like a place without stigma, that withdrawal may land as rejection.

Risk Was Named Inconsistently

The model also acknowledged risk unevenly depending on what was disclosed. On the first turn, the estimated probability that it named the risk was about 85 percent for prior non-suicidal self-injury, but only about 27 percent for prior suicidal ideation. Specific resources, such as a named crisis line, appeared in only 41 percent of responses overall, and were far more likely to follow a disclosure of a prior suicide attempt than an expression of hopelessness. A vulnerable person may reasonably hear that silence as minimization.

The Honest Counterweight

The model was not callous. It nearly always expressed empathy, and it nearly always encouraged the user to reach out to someone. Acknowledgment of risk actually improved as disclosures accumulated, rising about 62 percent every two turns. So this is not a story about a model that ignores distress. It is a narrower and stranger problem. The model hears you, says something kind, and then closes the door.

Why an Open-Source Model

We tested OLMo-2-32b, released by the Allen Institute for AI in April 2025 and competitive with GPT-4o mini on academic benchmarks. We chose it because it is fully open, including its training data, code, and analysis tools. Testing a closed commercial product can tell you that something happened but never why. Because OLMo-2-32b is open, other researchers can now take this finding and go looking for the mechanism inside the model itself.

What This Does Not Show

One model, and five of the 5,040 possible orderings of seven risk factors. We cannot say how every chatbot behaves. Our sample of therapists is also not the population most likely to turn to a chatbot in a crisis, so these results speak to alignment with clinical guidelines rather than to measured effects on real users. But a hazard found in some circumstances is a hazard that later work has to account for, and related research evaluating several commercial models found similarly inconsistent behavior, which suggests this is not unique to OLMo.

Why This Matters for Training

Policymakers are already moving. Illinois has prohibited AI from providing mental health services. Most of that debate has centered on chatbots doing something actively harmful, such as reinforcing delusions or assisting in the planning of a suicide attempt. This study adds a quieter item to the list. Withdrawal is also a failure, and it is much harder to see than a headline-generating harm.

For clinicians, the practical point is that your clients may be having these conversations already, and the model may be pulling back at exactly the moment a person needs it to lean in. That is why AI literacy is now part of how we train therapists, as described in our Statement on AI, and why it runs through the Sentio MFT program, our Clinical Supervisor Training, and our Free AI Course for Mental Health Professionals.

Read the full paper: "Independent Clinical Evaluation of General-Purpose LLM Responses to Signals of Suicide Risk" (Proceedings of IASEAI Conference, 2026, 2(1), 279-292). Supplementary materials are available on arXiv.

For more on Sentio's AI research and training, visit the Sentio AI Research Hub.