Overview

This narrative review synthesized findings from 11 studies (2021–2025) evaluating AI-driven chatbots across the suicide care continuum. A three-tier performance landscape emerges clearly: (1) clinically designed, evidence-anchored chatbots show feasibility and meaningful efficacy; (2) general-purpose commercial chatbots used for risk assessment show systematic intermediate-risk calibration failure and lethality information safety risks; and (3) commercially deployed mental health apps without clinical validation demonstrate near-universal crisis response inadequacy. The primary determinant of safe and effective chatbot performance in suicide care is clinical intentionality in design, not the underlying AI technology.

Chatbot performance in suicide risk assessment: Concordance and discordance with the literature

The finding that chatbots demonstrate adequate response alignment at the extremes of suicide risk but systematic failure at intermediate levels (McBain et al., Reference McBain, Cantor, Zhang, Baker, Burnett, Kofner, Breslau, Stein, Mehrotra and Yu2025) is consistent with the broader challenge of suicidal behavior prediction documented by Franklin et al. (Reference Franklin, Ribeiro, Fox, Bentley, Kleiman, Huang, Musacchio, Jaroszewski, Chang and Nock2017), who found that across 50 years of research that predictive accuracy across all methods remains only marginally above chance. (Belsher et al., Reference Belsher, Smolenski, Pruitt, Bush, Beech, Workman, Morgan, Evatt, Tucker and Skopp2019), in a systematic review of 17 machine learning suicide prediction models, found that while global classification accuracy was good (AUC ≥0.80 in most models), the positive predictive validity for suicide mortality was extremely low (PPV ≤0.01 in most models), such that a positive prediction result carried near-zero clinical utility, a structural limitation mirrored in the intermediate-risk failure observed across chatbots. Carter et al. (Reference Carter, Milner, McGill, Pirkis, Kapur and Spittal2017) found that even established clinical risk scales show pooled positive predictive values as low as 5.5%, contextualizing ChatGPT’s underestimation within a broader historical challenge.

The bidirectional miscalibration observed by Lauderdale et al. (Reference Lauderdale, Schmitt, Wuckovich, Dalal, Desai and Tomlinson2025) – underrating the most acute cases and overrating the least acute – extends beyond earlier single-direction underestimation findings (Elyoseph and Levkovich, Reference Elyoseph and Levkovich2023) and suggests that chatbots are applying population-level base-rate priors that compress risk estimates toward the mean, a regression-to-the-mean artifact. This is particularly dangerous at the acute high end of the risk spectrum: treatment planning by these chatbots was predicted by chronic but not acute risk ratings, whereas acute risk is the primary driver of emergency intervention decisions. GPT-4o’s recommendation of hospitalization for all veterans regardless of acuity level further illustrates an overcautious failure mode that would generate resource overutilization in real-world deployment. The finding by McBain et al. (Reference McBain, Cantor, Zhang, Baker, Burnett, Kofner, Breslau, Stein, Mehrotra and Yu2025) that intermediate-risk queries received no statistically different response rates than low-risk queries (despite expert consensus that they required different handling) confirms that current commercial chatbots lack the risk differentiation capability that is clinically essential for safe crisis management. Furthermore, the provision of direct answers to lethality-enabling questions about firearms and poisons by both ChatGPT and Claude represents a distinct safety risk beyond classification failure alone.

Thomas et al. (Reference Thomas, Elyoseph, Kuchinke and Meinlschmidt2025) provide complementary evidence from a non-chatbot context: using the Mixtral LLM to rate 100 real youth crisis helpline transcripts against NGASR expert ratings, they found that zero-shot prompting at temperature 0 yielded perfect inter-rating reliability (α = 1.00), while few-shot prompting showed best human-AI agreement for very high risk (α = 0.78). Critical clinical items showed poor validity regardless of configuration. Although Mixtral was employed as a backend classifier rather than a chatbot in this study, the finding that clinical item-level validity remains poor even under optimal LLM configurations has direct implications for the capabilities of deployed chatbots, whose underlying models share these structural limitations. Across all five included studies evaluating chatbot risk rating accuracy, no single chatbot model consistently matched clinician performance across all evaluated dimensions, and substantial inter-model variability was documented. This argues against the assumption that any commercially available chatbot can serve as a reliable clinical risk assessment tool without domain-specific clinical design.

Evaluating chatbot efficacy: Outcome heterogeneity and standards

A critical cross-cutting observation is the substantial heterogeneity in how chatbot effectiveness was operationalized across the 11 included studies: Likert-scale risk ratings against professional norms (Elyoseph and Levkovich, Reference Elyoseph and Levkovich2023; Levkovich and Elyoseph, Reference Levkovich and Elyoseph2023; Shinan-Altman et al., Reference Shinan-Altman, Elyoseph and Levkovich2024); VHA risk stratification alignment (Lauderdale et al., Reference Lauderdale, Schmitt, Wuckovich, Dalal, Desai and Tomlinson2025); direct response likelihood categorized by risk level (McBain et al., Reference McBain, Cantor, Zhang, Baker, Burnett, Kofner, Breslau, Stein, Mehrotra and Yu2025); pre-specified adequacy criteria (Pichowicz et al., Reference Pichowicz, Kotas and Piotrowski2025); evidence-based care delivery rates (Dimeff et al., Reference Dimeff, Jobes, Koerner, Kako, Jerome, Kelley-Brimer, Boudreaux, Beadnell, Goering, Witterholt, Melin, Samike and Schak2021); multi-domain efficacy ratings (Cui et al., Reference Cui, Gu, Fang and Zhu2025); self-reported ideation cessation (Maples et al., Reference Maples, Cerit, Vishwanath and Pea2024); and content quality coding (Martinengo et al., Reference Martinengo, Lum and Car2022). This eight-framework heterogeneity across 11 studies precludes quantitative synthesis and reflects the absence of a consensus minimum performance standard for chatbot suicide assessment tools. The field urgently requires a standardized benchmark framework analogous to the C-SSRS (Posner et al., Reference Posner, Brown, Stanley, Brent, Yershova, Oquendo, Currier, Melvin, Greenhill, Shen and Mann2011), adapted for evaluating chatbot performance across the suicide care continuum.

Chatbot-based clinical interventions and the design–efficacy relationship

Martinengo et al. (Reference Martinengo, Lum and Car2022) demonstrated adequate CBT exercise delivery but inadequate crisis response in nine commercial chatbots. Firth et al.’s (Reference Firth, Torous, Nicholas, Carney, Pratap, Rosenbaum and Sarris2017) meta-analysis of 18 smartphone intervention RCTs (N = 3,414) reported significant but modest reductions in depressive symptoms (Hedges’ g = 0.38), contextualizing the therapeutic effect sizes observed across chatbot studies. Critically, a meta-analysis of 19 RCTs of standalone mental health apps found no significant pooled effect on suicidal ideation or self-injury (Hedges’ g ranging from −0.14 to 0.18 across these outcomes), even while modest effects on depression were observed (g = 0.33), underscoring the difficulty of achieving suicidality-specific outcomes through unguided digital tools and reinforcing why clinical framework integration is essential (Weisel et al., Reference Weisel, Fuhrmann, Berking, Baumeister, Cuijpers and Ebert2019). The design–efficacy relationship is now clearly established across the 11 included studies. Chatbots explicitly designed around clinical frameworks, including Jaspr Health (CAMS model; (Dimeff et al., Reference Dimeff, Jobes, Koerner, Kako, Jerome, Kelley-Brimer, Boudreaux, Beadnell, Goering, Witterholt, Melin, Samike and Schak2021) and the (Cui et al., Reference Cui, Gu, Fang and Zhu2025) chatbot (ACT framework), both demonstrated high efficacy outcomes in their respective evaluation contexts. Chatbots deployed without clinical framework grounding, including the 29 commercial apps in Pichowicz et al. (Reference Pichowicz, Kotas and Piotrowski2025) and the 9 apps in Martinengo et al. (Reference Martinengo, Lum and Car2022) demonstrated inadequate or moderate performance. This dose–response relationship between clinical design and chatbot efficacy is the strongest practical conclusion of this review. The key differentiator between clinically anchored and commercially deployed chatbots is clinical design: ACT framework-anchored prompt engineering, as demonstrated by Cui et al. (Reference Cui, Gu, Fang and Zhu2025), produced substantially better crisis-relevant performance than general-purpose chatbot deployment, while Jaspr Health was purpose-built around the CAMS theoretical framework with clinical validation, in contrast to the 29 apps evaluated by Pichowicz et al. (Reference Pichowicz, Kotas and Piotrowski2025), which were general mental distress apps without crisis-specific clinical validation.

Moylan and Doherty’s (Reference Moylan and Doherty2025) expert analysis of Wysa and Replika from a mental health professional perspective confirms this from the clinician side: professionals identified generic modes of care, dependency-promoting interaction patterns, and harm-inviting response designs as primary concerns, underscoring that clinical practitioner perspectives should be systematically integrated into chatbot development processes.

Real-world deployment and population-level reach

Maples et al. (Reference Maples, Cerit, Vishwanath and Pea2024) found that 3% of Replika users reported halted suicidal ideation. These findings suggest that chatbots may operate primarily through indirect pathways, alleviating loneliness and supporting emotional regulation, rather than direct crisis intervention, consistent with Fitzpatrick et al.’s (Reference Fitzpatrick, Darcy and Vierhile2017) RCT of Woebot demonstrating moderate and significant depression symptom reductions (d = 0.44) at 2 weeks. Qualitative evidence from 19 real-world users of generative AI chatbots for mental health corroborates both the potential and the limitations identified in clinical evaluations: users reported meaningful benefits, including a sense of emotional sanctuary and personal insight into relationships, while simultaneously identifying the urgent need for stronger safety guardrails and more persistent clinical integration (Siddals et al., Reference Siddals, Torous and Coxon2024). Pichowicz et al.’s (Reference Pichowicz, Kotas and Piotrowski2025) finding that none of the 29 commercially available chatbot agents provided an adequate crisis response contextualizes these modest benefits within a larger landscape of inadequate real-world deployment. The contrast between Dimeff et al.’s (Reference Dimeff, Jobes, Koerner, Kako, Jerome, Kelley-Brimer, Boudreaux, Beadnell, Goering, Witterholt, Melin, Samike and Schak2021) findings and those of Pichowicz et al. (Reference Pichowicz, Kotas and Piotrowski2025) is central to understanding the chatbot performance landscape; it confirms that chatbot efficacy is not a property of the underlying technology but rather of clinical intentionality in design. This inadequacy has historical precedent; a 2016 systematic evaluation of 49 interactive suicide prevention apps found that only 26.5% contained at least one WHO-endorsed evidence-based strategy, none provided comprehensive evidence-based support, and potentially harmful content was identified in a subset of apps (Larsen et al., Reference Larsen, Nicholas and Christensen2016), suggesting that the commercial app ecosystem has made limited progress in nearly a decade. The implication is that the 3% direct benefit found by Maples et al. (Reference Maples, Cerit, Vishwanath and Pea2024) may represent an upper bound for commercially deployed chatbots lacking clinical validation, not a baseline.

Post-crisis monitoring and system integration

The SIMON protocol (Sels et al., Reference Sels, Homan, Ries, Santhanam, Scheerer, Colla, Vetter, Seifritz, Galatzer-Levy, Kowatsch, Scholz and Kleim2021) addresses the post-discharge high-risk transition documented by Chung et al. (Reference Chung, Hadzi-Pavlovic, Wang, Swaraj, Olfson and Large2019), who found pooled suicide rates of 2,950 per 100,000 person-years in the first week post-discharge, confirming that this period represents the highest-risk window in psychiatric care. Episodic routine follow-up represents a primary limitation of current post-discharge care, and this is precisely the gap that SIMON’s continuous chatbot monitoring is designed to address. Cui et al.’s (Reference Cui, Gu, Fang and Zhu2025) self-help platform represents a scalable complement to SIMON’s monitoring model. The (Cui et al., Reference Cui, Gu, Fang and Zhu2025) self-help web chatbot platform offers a complementary model for post-crisis engagement, providing accessible intervention delivery that does not depend on clinical infrastructure; this is a critical advantage given McBain et al.’s (Reference McBain, Cantor, Zhang, Baker, Burnett, Kofner, Breslau, Stein, Mehrotra and Yu2025) observation that the United States has approximately one psychiatrist per 13,492 residents, addressing the scalability limitations that have constrained prior monitoring approaches. Of the 11 included studies, only Dimeff et al. (Reference Dimeff, Jobes, Koerner, Kako, Jerome, Kelley-Brimer, Boudreaux, Beadnell, Goering, Witterholt, Melin, Samike and Schak2021) designed a chatbot for integration within an existing clinical workflow, and Pichowicz et al. (Reference Pichowicz, Kotas and Piotrowski2025) confirmed that crisis service linkage failures are the most common deficiency across commercially deployed apps. The disconnect between technology development and crisis service infrastructure is clearly reflected in the current evidence base. This pattern is consistent with previously identified fundamental misconceptions in how digital mental health technologies are conceptualized, developed, and integrated into care systems, including treating technology as a finished product rather than a service, and paying insufficient attention to user engagement and clinical workflow integration (Mohr et al., Reference Mohr, Weingardt, Reddy and Schueller2017).

Toward clinically grounded chatbot development: Recommendations for training and design

A critical gap across the included studies is the absence of a clinical framework grounding in most evaluated chatbots. The systematic calibration failures observed across risk assessment studies reflect training data skewed toward general text corpora, where extreme-risk presentations are statistically rare, producing regression-to-the-mean errors at precisely the clinically critical high-risk end. Cui et al.’s (Reference Cui, Gu, Fang and Zhu2025) demonstration that GPT-4, when fine-tuned through structured ACT framework prompt engineering, achieved high intervention quality without new clinical training data, provides an immediately actionable pathway: clinical framework-guided prompt engineering can substantially improve chatbot performance in the near term while clinically curated training datasets are developed. Future chatbot training should incorporate: (1) curated clinical datasets derived from validated structured clinical instruments (C-SSRS, CAMS) with deliberate oversampling of high-risk presentations; (2) passive biometric sensing and ecological momentary assessment combined with structured clinical data (Bourla et al., Reference Bourla, Mouchabac, El Hage and Ferreri2018; Sels et al., Reference Sels, Homan, Ries, Santhanam, Scheerer, Colla, Vetter, Seifritz, Galatzer-Levy, Kowatsch, Scholz and Kleim2021); and (3) culturally diverse training data. All 11 included studies are from high-income countries, with only (Cui et al., Reference Cui, Gu, Fang and Zhu2025) representing a non-Western context.

Potential harms, ethical considerations, and governance

The ethical risk profile of chatbots in suicide care is now two-sided. McBain et al. (Reference McBain, Cantor, Zhang, Baker, Burnett, Kofner, Breslau, Stein, Mehrotra and Yu2025) established that current commercial chatbots can provide lethality-enabling information under some query conditions, while Pichowicz et al. (Reference Pichowicz, Kotas and Piotrowski2025) established that they simultaneously fail to provide adequate crisis redirection. Emerging from 9,000 standardized responses, the McBain et al. (Reference McBain, Cantor, Zhang, Baker, Burnett, Kofner, Breslau, Stein, Mehrotra and Yu2025) findings establish that current commercial chatbots can facilitate harm not merely through inadequate crisis response but through the active provision of lethality-enabling information. This bidirectional risk – potential harm facilitation and inadequate harm prevention – has no analog in prior human clinical practice and is not addressed by any currently existing regulatory framework for AI mental health tools. Moylan and Doherty’s (Reference Moylan and Doherty2025) finding that mental health professionals identified medium-to-low trust in commercial chatbots and multiple harm vectors (generic care, dependency risk, and manipulation potential) reinforces the urgent need for domain-specific governance frameworks developed in partnership with clinical experts. Prior evidence indicates that depression severity is consistently associated with problematic smartphone use, with at least medium effect sizes across reviewed studies (Elhai et al., Reference Elhai, Dvorak, Levine and Hall2017), suggesting a pre-existing vulnerability profile among chatbot users that amplifies dependency risks. The (Herbener and Damholdt, Reference Herbener and Damholdt2025) cross-sectional study of 1,599 Danish high school students corroborates the dependency risk profile identified by Maples et al. (Reference Maples, Cerit, Vishwanath and Pea2024), demonstrating that socially disconnected young people are disproportionately drawn to chatbots as emotional coping tools; social-supportive chatbot users reported significantly more loneliness (d = 0.53) and less perceived social support (d = 0.46) than non-chatbot users.

Campbell et al.’s (Reference Campbell, Babb, Lambie and Hayes2025) longitudinal content analysis of seven chatbots across two phases (spring 2023 and summer 2024) found measurable improvement in the depth and accuracy of suicide-related responses over 1 year, with increased emphasis on the 988 crisis hotline and more comprehensive information about suicide warning signs. However, Microsoft Copilot provided an unsolicited image depicting suicide in Phase 1, illustrating that even improving AI systems remain capable of harmful outputs. This temporal evidence suggests that chatbot safety in suicide contexts is improving but remains uneven, and requires periodic independent review rather than one-time evaluation.

Limitations

This narrative review has several limitations. The 11 included studies are exclusively from high-income countries, limiting generalizability to low- and middle-income countries and non-Western populations. Cross-cultural research has demonstrated that suicidal behavior and its communicative function differ meaningfully between Western and non-Western contexts. For instance, the four-factor semiotic model empirically supported in Western populations was not applicable in Uganda, where suicidal behavior followed a distinct two-factor communicative structure (Hjelmeland et al., Reference Hjelmeland, Knizek, Kinyanda, Musisi, Nordvik and Svarva2008), suggesting that chatbot systems trained on Western clinical data may not perform equitably across cultural settings. The narrative review design carries inherent risks of selection bias. Outcome heterogeneity precluded quantitative synthesis. Publication bias cannot be excluded.