Artificial Intelligence in suicide prevention: promise, peril, and clinical implications for clinical practice

Abstract

The integration of Artificial Intelligence (AI) into psychiatry is progressing rapidly, with promising yet underexplored applications in suicide prevention. While most existing studies have focused on suicide risk prediction, AI could be applied more broadly across the suicide prevention continuum. This review explores AI’s potential in improving the detection of suicidal ideation (SI), enhancing clinician training, and expanding patient access to psychotherapeutic support. Studies suggest that AI models may identify SI more sensitively than clinicians by detecting subtle linguistic or semantic cues. Furthermore, AI-based virtual patients have been shown to improve clinicians’ confidence and accuracy in suicide risk assessments, particularly when used alongside traditional training methods. On the patient side, CBT-inspired AI tools offer scalable, on-demand support, but most remain unregulated and clinically unvalidated. Despite their perceived empathy and availability, such tools pose ethical and clinical risks. Many patients use AI tools without professional guidance, a trend clinicians often underestimate. This disconnect calls for urgent regulatory oversight and professional engagement. Without clear clinical validation, AI-based tools may evolve outside medical frameworks, exposing vulnerable individuals to preventable harm. Psychiatry must not only adapt but also lead, ensuring that innovation is harnessed responsibly to serve patients with empathy, safety, and scientific rigor.

Introduction

The use of Artificial Intelligence (AI) is spreading rapidly across developed countries, including among patients with mental disorders. In clinical neuroscience, AI can be viewed as a set of computational tools that mimic certain aspects of human cognition—such as information processing, language comprehension, and risk evaluation. When applied to psychiatry, AI models can process and interpret complex data to identify mental health risks, assist in clinical diagnosis, or personalise interventions. Importantly, AI does not replicate human understanding or empathy, but it can detect subtle markers in data that may escape human attention, such as linguistic patterns associated with suicidal ideation.

In a recent U.S. survey among adults with a mental disorder who had used an AI large language model (LLM), such as ChatGPT, within the past year, nearly half reported using it as a form of psychological support (Rubin 2026). Even more strikingly, two-thirds of these individuals stated that interacting with AI improved their mental health condition. In younger populations, AI use for mental health purposes is also substantial. A recent survey of adolescents and young adults (ages 12–21) found that approximately 13% used AI for mental health advice, with this percentage increasing significantly among those over 18. Among AI users, over 65% used it monthly for mental health-related advice, and more than 90% found it very helpful (McBain et al. 2025). These numbers highlight that a significant proportion of individuals with mental health conditions are using AI tools without any clinical supervision. Alarmingly, most psychiatrists are unaware of this widespread usage—or outright deny it—despite its potential impact on both patients and clinical practice (Frances, 2026). When used correctly, AI could offer substantial benefits in mental health care. However, unsupervised, or inappropriate use may lead to serious and even life-threatening consequences.

This issue is particularly critical in the field of suicidal behaviours (i.e., suicide and suicide attempts). A recent tragic case involved a suicidal adolescent who disclosed his suicidal thoughts to an AI-based LLM and eventually died by hanging, reportedly after the AI explained to him how to tie the knot (Nagata et al. 2026). Despite such catastrophic events, AI may still hold promise in improving care for suicidal individuals (Lee et al. 2021). Numerous studies have shown promising results in predicting suicide risk using LLMs and machine learning (ML) algorithms, based on electronic health records or social media data (Arowosegbe and Oyelade 2023; Levkovich and Omar 2024; Badian et al. 2023; Mansoor and Ansari 2024; Du et al. 2018; Wu et al. 2023; Sherekar and Mehta 2025). While improving the prediction of suicidal behaviour is undeniably important, many other potential applications of AI in suicide prevention remain underexplored. Indeed, the implementation of predictive models in real-life clinical practice is still far from reality. In contrast, some AI-based tools could be applied more immediately to support suicide prevention, such as: (i) detecting suicidal ideation in clinical settings through analysis of clinical interviews; (ii) improving clinician training; and (iii) supporting patients in coping with suicidal thoughts. While these applications may offer real benefits, it is essential to consider the potential risks of relying too heavily on AI, especially in the context of vulnerable individuals. Accordingly, the aim of this review is to examine these three domains of direct AI application in suicide prevention, while critically discussing their potential benefits and associated risks and to address the broader challenges of integrating AI safely into suicide prevention practices.

Clinical judgment and Artificial Intelligence: Complementarity, not substitution

Suicidal behaviours can be conceptualised as existing along a continuum, ranging from passive suicidal ideation (SI) to active ideation, suicide attempts, and ultimately death by suicide. Identifying patients experiencing SI is thus the crucial first step in this chain, offering an opportunity to intervene early and prevent progression to more severe outcomes (Nobile et al. 2022; Courtet and Baca-García 2024). This is not only essential from a suicide prevention standpoint, but also because SI is itself associated with worse psychiatric prognosis, including treatment resistance and recurrence in mood disorders (Nobile et al. 2024; Bergfeld et al. 2018).

Yet, detecting SI remains a substantial clinical challenge. Studies have revealed a concerning level of discordance—up to 75%—between clinicians’ evaluations and patients’ self-reports regarding suicidal thoughts (Nobile et al. 2023; Nobile et al. 2024; Gao et al. 2015). In most cases, patients reported SI that clinicians failed to detect. However, this discrepancy cannot be attributed solely to limitations in clinical judgement. A significant body of evidence suggests that some patients, particularly adolescents and individuals from diverse or vulnerable backgrounds, may not disclose suicidal thoughts during clinical encounters due to stigma, fear of hospitalisation, uncertainty, or a weak therapeutic alliance (Uhl et al. 2023). This is far from a trivial oversight: approximately 56% of individuals who die by suicide have consulted a healthcare professional within the four weeks preceding their death (O’Rourke et al. 2023). Moreover, patients who only endorse SI via self-report questionnaires, but not during clinical interviews, appear to carry a suicide attempt risk equivalent to those whose SI is confirmed by clinicians (Nobile et al. 2024; Uhl et al. 2023). These findings highlight both the limitations of clinical judgement alone and the challenges related to patient disclosure underscoring the need for enhanced tools to detect SI more reliably.

AI, and in particular large LLMs, could be instrumental in this regard. Some studies have begun to explore the capacity of AI to assess presence of SI based on patient narratives or structured data, offering an alternative and potentially more sensitive detection method. For instance, a recent study compared SI assessments made by ChatGPT-4 and ChatGPT-3.5 against those of 379 mental health professionals on patient’s vignettes (Levkovich and Elyoseph 2023). Their findings suggest that ChatGPT-4 matched clinicians’ accuracy in predicting suicide attempts and even showed greater sensitivity in detecting SI and psychological pain. In contrast, ChatGPT-3.5 significantly underestimated suicide risk across all conditions tested, particularly in the most severe cases. These results highlight two key points. First, AI systems, particularly more recent versions, may be capable of identifying SI with a level of precision comparable to human clinicians. Second, different AI models vary widely in their performance, and under certain conditions, older versions may dangerously underestimate risk. This calls for caution: while AI may enhance SI detection, it also introduces the possibility of error, particularly when used uncritically or without awareness of model limitations (Arowosegbe and Oyelade 2023; Levkovich and Omar 2024; Bernert et al. 2020).

Importantly, AI could also serve as a valuable complement in clinical settings, by offering a form of ‘second reading’ of clinical interviews. One could imagine a system that transcribes patient–clinician dialogues in real-time and analyzes subtle linguistic markers suggestive of SI (e.g., increased use of intensifiers or personal pronouns, or tone shifts associated with hopelessness). Such markers may escape clinician attention, especially in time-constrained or emotionally charged consultations. Moreover, such approaches may be particularly valuable in identifying patients who do not explicitly disclose suicidal thoughts, by detecting indirect or implicit linguistic signals associated with distress or suicidality. In this regard, a very recent study demonstrated that AI models trained on naturalistic patient–provider interviews were able to detect SI using both acoustic and semantic features, providing additional cues that may not be consciously perceived by clinicians (Lho et al. 2025).

Nevertheless, these tools should be used with great caution. Over-reliance on AI could lead to automation bias, where clinicians overly trust algorithmic outputs even when these conflict with their own judgement or the clinical context (Perlis 2026). In practice, these kinds of misjudgments could either inflate perceived risk, leading to unnecessary interventions, or deflate it, missing patients in need of urgent care. Clinicians must therefore remain the primary decision-makers. To avoid inappropriate use and maximise AI’s potential, clinical training must integrate AI literacy, especially in suicide prevention (Elyoseph et al. 2024). Moreover, beyond immediate decision-making risks, an additional concern relates to the long-term impact of AI integration on the development of clinical expertise. As AI tools become increasingly embedded in training and practice environments, future clinicians may rely on them from the earliest stages of their education. This raises the possibility of cognitive offloading, where critical aspects of clinical reasoning are delegated to AI systems, potentially limiting the development of independent judgement and expertise. Indeed, emerging evidence suggests that increased reliance on AI tools may be associated with reduced critical thinking abilities, with cognitive offloading acting as a mediating factor (Gerlich 2025). Over time, such reliance could lead to a form of ‘AI learning dependency’, in which clinicians may struggle to critically evaluate or challenge algorithmic outputs, particularly in complex or ambiguous clinical situations such as suicidality. From an educational perspective, this also raises the question of when AI should be introduced during clinical training. Premature exposure to AI tools may hinder the development of independent clinical reasoning, as trainees may rely on algorithmic support before acquiring core assessment skills. A staged approach could therefore be considered, in which AI is introduced later in training once foundational competencies in clinical judgement have been established. Such an approach may help ensure that AI enhances, rather than supplants, the development of critical thinking and decision-making skills in future clinicians.

In sum, while AI may improve the detection of SI, particularly in cases where clinicians might miss subtle cues, it should not be viewed as a substitute for clinical expertise. Used in a complementary manner, AI has the potential to bridge current gaps in SI detection and contribute meaningfully to prevention strategies.

Training the human copilot: AI as a tool for clinician education in suicidal behaviors

Given the high rates of under-detection of SI by clinicians and the potential support AI can offer in refining risk assessments, it is crucial to integrate AI technologies not only into clinical tools but also into the training of healthcare professionals. Importantly, this applies not only to psychiatrists but also to general practitioners (GPs), who are frequently the first medical professionals to encounter patients with suicidal thoughts (Stene-Larsen and Reneflot 2019; Feldman et al. 2007). Even though mood disorders are a primary risk factor for suicidal behaviour, fewer than 40% of GPs report routinely evaluating SI in patients with mood disorders (Feldman et al. 2007). Thus, strengthening the ability of psychiatrists but also GPs and other non-specialist clinicians to detect SI is a pressing public health objective.

In this context, the integration of AI-based tools into clinician training offers a promising path forward. These tools can serve dual purposes: first, by educating practitioners on how AI systems work, what their limitations are, and how to interpret their outputs cautiously; and second, by actively assisting in the acquisition of clinical interviewing and risk assessment skills. One of the most innovative uses of AI in this domain is the deployment of virtual patient simulators, chatbots powered by natural language processing, which allow clinicians to practice suicide risk assessments in a controlled environment. For example, a virtual agent named ‘Alex’ was developed and designed to train mental health clinicians in assessing suicide risk (Foster et al. 2015). This virtual interaction provided a safe and repeatable learning experience. Clinicians trained with Alex were found to be more confident and effective in eliciting information related to suicidal thoughts compared to those trained using traditional methods. Similarly, another study evaluated the impact of a conversational agent designed to simulate patients presenting with SI (Martínez-Miranda 2017). The study demonstrated that clinicians trained with this AI-powered agent showed better recognition of verbal and non-verbal cues associated with suicide risk, and greater ease in navigating difficult conversations than clinicians trained with traditional methods. These findings support the use of AI-based simulations as a scalable and engaging complement to classical educational formats such as workshops or lectures. More recently, a study investigated the use of an AI-driven virtual reality training program targeted at young clinicians from 20 different countries, each interacting in their native language, demonstrating the potential of such tools to be deployed across linguistic and cultural contexts (Elyoseph et al. 2026). This tool incorporated real-time feedback and naturalistic scenarios, helping trainees fine-tune their ability to assess SI and behaviours. The results showed significant improvements in both knowledge acquisition and clinical confidence, particularly in detecting SI in ambiguous or subtle presentations. Together, these studies suggest that AI can be effectively leveraged to standardise and enhance training across a wide range of healthcare professionals, including those who may not have extensive psychiatric expertise (Yıldız 2025).

Traditional training in suicide risk assessment relies heavily on theoretical instruction combined with supervised role-playing exercises. While these methods are essential for developing interviewing skills and clinical confidence, they are also time-consuming and resource intensive. Moreover, opportunities for repeated role-play practice are necessarily limited by institutional constraints, availability of supervisors, and logistical factors. As a result, many clinicians have only intermittent access to realistic training situations, which may hinder long-term skill consolidation. In this context, AI-based training tools offer a valuable complement to conventional education by providing continuous, on-demand opportunities for practice. After completing standard training programs, clinicians could use AI-powered virtual patients to rehearse assessments, refine their questioning strategies, and receive feedback at their own pace. Such an approach would promote sustained learning and professional development, allowing practitioners to maintain and strengthen their suicide risk assessment skills over time. However, the implementation of AI-based training tools raises important practical and institutional questions that remain largely unaddressed. First, the development of such tools should not be left solely to technology companies but should involve close collaboration between clinicians, researchers, and professors to ensure clinical relevance, safety, and pedagogical validity. Second, these tools should be developed and evaluated within rigorous frameworks, similar to other clinical or educational interventions, including external validation and oversight by academic and regulatory bodies. Nevertheless, it is important to acknowledge that systematically training clinicians in AI literacy remains a significant challenge, as busy clinical environments, competing demands, and limited institutional resources often hinder the implementation of large-scale, sustained educational initiatives. Moreover, the adoption of AI-based training tools will critically depend on clinicians’ acceptability, which remains uncertain. Emerging evidence suggests that while clinicians are generally open to AI-supported decision-making, this acceptance is tempered by significant concerns, including liability, increased workload, alert fatigue, and limited resources within healthcare systems (Isnard-Bagnis et al. 2025; Bentley et al. 2022). Moreover, digital tools may contribute to cognitive overload, stress, and burnout if not carefully designed and implemented. Acceptability is therefore not solely determined by technological performance, but also by perceived usefulness, ease of use, trust, and integration into clinical workflows (Isnard-Bagnis et al. 2025). These findings highlight the importance of involving clinicians in the co-design of AI tools and ensuring that their implementation is accompanied by adequate training, institutional support, and clear clinical frameworks.

In summary, the inclusion of AI in clinician education serves a dual function: increasing awareness about how AI operates in the clinical setting and offering innovative platforms for learning critical suicide risk assessment skills. Given the persistent gaps in SI detection such tools may represent an invaluable asset in equipping the healthcare workforce to better identify and respond to suicidal patients.

When patients turn to AI alone: Uses, perceptions, and risks of unsupervised mental health support

Cognitive Behavioural Therapies (CBTs) specifically designed to target SI have demonstrated strong efficacy across various clinical trials. Interventions such as Cognitive Therapy for Suicide Prevention (CT-SP) or Brief Cognitive Behavioural Therapy for Suicide Prevention (BCBT-SP) have significantly reduced SI and suicide attempts in high-risk populations (Ducasse et al. 2018; Bryan et al. 2018; Swift et al. 2021). However, access to such therapies remains limited due to a global shortage of trained therapists, long waitlists, and uneven geographical distribution of services. In this context, AI-based tools—particularly those leveraging natural language processing (NLP) to deliver structured, CBT-inspired interventions—have emerged as promising candidates to help close the accessibility gap. These tools promise on-demand, scalable support that can be deployed across diverse populations. However, most existing AI-driven mental health applications are not regulated or clinically validated. To bypass the need for medical device approval from agencies such as the U.S Food and Drug Administration (FDA) or European Medicines Agency (EMA), developers often label them as ‘wellness apps’, effectively excluding them from rigorous clinical oversight (Rubin 2026; Frances, 2026). This strategic framing raises profound concerns, particularly when such tools are used in high-risk contexts such as suicide prevention.

Users often engage with AI chatbots in moments of psychological distress, sometimes discussing suicidal thoughts or urges. Without proper validation, such tools may provide a false sense of safety or inadvertently deliver harmful advice (Rubin 2026). Beyond technical efficacy, the appeal of AI-based mental health tools lies partly in their perceived emotional accessibility. Indeed, patients frequently described AI tools as more empathic, less judgmental, and more available than human therapists (Rubin 2026; Bickmore et al. 2010). Particularly for stigmatised conditions such as suicidal behaviours, interacting with an AI may lower the threshold for disclosure. This phenomenon may partially explain the widespread, unmonitored use of LLMs like ChatGPT for psychological support. However, this seemingly non-judgmental stance can also lead to sycophantic behaviour, where the AI aligns with the user’s perspective, even when that perspective reflects distorted or dangerous thinking. This tendency was observed in recent studies, warning that LLMs may validate SI without offering appropriate redirection or intervention (Rubin 2026; Hudon and Stip 2025). In such cases, the very quality that makes AI appealing to users, its unconditional agreement, may also prevent critical therapeutic interventions.

The risks are not only hypothetical. As highlighted by Hudon et al. (Hudon and Stip 2025), AI tools may inadvertently reinforce avoidance behaviours, allowing users to sidestep real-life help-seeking such as contacting emergency services, family members, or clinicians. Some users in acute distress have engaged extensively with AI agents without ever reaching out to a human professional. This is especially concerning given that most LLMs are not designed to manage psychiatric crises, assess imminent risk, or make real-time safety decisions. Hudon et al. (Hudon and Stip 2025), also articulate a broader theoretical concern: that the use of AI in psychotherapy may generate an ‘illusion of clinical interaction’. By simulating empathic dialogue without genuine intersubjectivity, AI tools can create a pseudo-therapeutic space where the patient feels heard but is not actually held within a therapeutic alliance. They argue that this illusion could erode core aspects of clinical care, such as mutual responsibility, embodied presence, and ethical containment, replacing them with algorithmic simulations that are not equipped to manage the complexity of human suffering. The risk is compounded by the lack of accountability mechanisms. Indeed, many AI applications currently in use for mental health support are neither transparent in their algorithms nor consistent in their clinical responses (Rubin 2026). While some offer disclaimers about being non-medical tools, these are often ignored or misunderstood by users, particularly in moments of crisis. Moreover, the 24/7 availability of such tools, often touted as a major advantage, may paradoxically enable chronic self-isolation or dependency on non-human agents for emotional regulation. AI tools can also misfire in more subtle ways. While some studies demonstrated that AI conversational agents can effectively engage users and build rapport, especially among patients with low health literacy, they did not assess the agents’ performance in high-risk psychiatric contexts like suicidality (Bickmore et al. 2010). In contrast, some other studies highlight cases where AI agents failed to detect dissociation, veiled references to self-harm, or escalating risk signals (Rubin 2026; Hudon and Stip 2025; Arendt et al. 2023; Pichowicz et al. 2025). The problem, then, is not simply that AI lacks empathy, but that it lacks the clinical discernment to act on nuanced cues, and the legal responsibility to do so. Despite these limitations, some applications show promise. For instance, an AI-powered agent was trained to interact with patients suffering from depression (Martínez-Miranda 2017). The system was well received, and some patients reported improvements in their understanding of suicidal risk and coping mechanisms. However, the authors caution that the agent was deployed in a tightly controlled environment, and its extrapolation to real-world settings remains uncertain.

More broadly, the widespread patient use of AI-based tools outside clinical oversight reflects a disconnect between psychiatric practice and evolving digital behaviours. As Frances emphasises (Frances, 2026), many mental health professionals remain surprisingly passive or even dismissive regarding the growing role of AI in mental health. The author attributes this inertia to ignorance, denial, and professional arrogance, summed up in the common refrain: ‘no machine will ever replace me’. This cognitive and institutional rigidity risks widening the gap between care provision and actual patient behaviours. If clinicians continue to ignore the increasing reliance on AI by their patients, they may not only miss an opportunity to guide its use responsibly but may also lose their role as central actors in the therapeutic process. At worst, this could erode trust, reinforce paternalistic models of care, and lead to preventable harms. Frances warns that if clinicians do not learn to work with AI tools, they may simply be replaced by them. Importantly, he argues that mental health professionals must refocus their training on areas where AI cannot yet reach, such as severe or complex psychiatric cases, and develop new competencies, including oversight of AI-based interventions and critical evaluation of their outputs.

From a regulatory and scientific standpoint, an urgent issue remains: no randomised controlled trials (RCTs) to date have tested AI-based applications as validated tools for suicide-specific CBT. This represents a significant gap in the literature. Given the high stakes involved in suicide prevention, the absence of validated tools is not only problematic but ethically concerning. Future research should prioritise the development and empirical validation of AI-based interventions that specifically target SI and behaviours, using well-established psychotherapeutic models, as it starts to be done in major depressive disorder (e.g., Woebot) (Fitzpatrick et al. 2017; Inkster et al. 2018). Beyond research, there is an urgent need for regulatory bodies (e.g., FDA, EMA) to take a more active role in overseeing the rapidly expanding ecosystem of AI-based mental health tools. Clear regulatory frameworks are needed to ensure minimum safety standards, transparency, and accountability. This could include mandating built-in safety mechanisms (e.g., automatic detection of suicide-related content, systematic redirection to crisis resources such as national suicide hotlines), as well as clearer public communication regarding the non-clinical nature of general-purpose AI systems such as ChatGPT or Gemini and others. Large-scale public health campaigns may be necessary to raise awareness about the limitations and risks of these tools when used for mental health support. In parallel, greater responsibility should be placed on developers to design AI systems that are safe by design, particularly in high-risk contexts such as suicidality. This includes implementing robust guardrails to prevent harmful or overly validating responses and ensuring that AI systems do not inadvertently reinforce suicidal thinking or discourage help-seeking behaviours. Waiting for these applications, clinicians must act as informed intermediaries, educating patients, evaluating tools, and integrating technology into care pathways in a manner that prioritises safety, transparency, and human oversight. The profession cannot afford to remain on the sidelines. In the coming years, the question will not be whether patients use AI for mental health, but whether psychiatry is ready to meet them where they already are.

In sum, AI has the potential to enhance access to suicide-focused psychotherapies, but only under strict regulatory, ethical, and clinical oversight. If AI is to support—not supplant—therapeutic care, developers and clinicians must work together to establish clear boundaries, accountability structures, and patient protections. As with any powerful tool, the question is not whether AI should be used, but how, when, and under whose guidance.

Ethical and cognitive challenges in integrating AI into suicide prevention

The integration of AI into suicide prevention raises not only technical or clinical questions but also profound ethical and cognitive dilemmas. While these issues are often framed around familiar concerns, such as data privacy, algorithmic bias, and transparency, suicide prevention introduces unique challenges, due to the vulnerability and acuity of the target population.

AI tools designed to predict suicidal behaviour often rely on massive datasets, ranging from electronic health records to social media activity. Yet, the ethical implications of these predictive systems are profound: how should we act on predictions that flag someone as high-risk? What thresholds justify intervention, especially when false positives can lead to coercive measures or increased stigma? Ethical frameworks must evolve to account for the non-binary, probabilistic nature of AI predictions, which are far from the deterministic decisions clinicians are accustomed to (Poudel et al. 2025).

One under-discussed issue is the ‘delegation dilemma’, the cognitive risk that clinicians may begin to offload judgement to algorithmic tools, particularly in high-pressure environments (Mörch et al. 2020). While AI can assist in processing large volumes of data, its integration into clinical reasoning must be carefully monitored. There is a growing risk of automation bias, where clinicians override their own instincts in favour of model outputs, even when these contradict patient narratives or clinical intuition. This is especially dangerous in suicide prevention, where context, nuance, and subtle interpersonal cues often guide decision-making. Beyond cognitive biases, concerns about legal liability may further complicate clinicians’ interactions with AI-based risk predictions. Qualitative studies have shown that clinicians may fear being held responsible if they disregard a high-risk prediction and an adverse outcome occurs, potentially leading to more defensive clinical decisions such as unnecessary hospitalisation (Bentley et al. 2022). Addressing these concerns will require the development of clear clinical guidelines, shared decision-making frameworks, and institutional policies that explicitly define the role of AI as a support tool rather than a determinant of clinical responsibility.

Patients increasingly perceive AI tools as more empathic than clinicians (Martínez-Miranda 2017). Yet this simulated empathy may be misleading since LLMs are sycophantic by design (Rubin 2026). In suicide prevention, this can lead to missed opportunities for therapeutic confrontation or safety planning. The illusion of being ‘heard’ may pacify distress temporarily but fails to engage in the complex relational work that underpins recovery.

Many AI applications collect sensitive psychological data, language samples, emotional states, behavioural patterns, yet often do so outside traditional medical settings, with vague or absent consent processes (Elyoseph et al. 2024). The regulatory vacuum surrounding mental health AI tools allows companies to avoid FDA or EMA scrutiny. This undermines both safety and trust, particularly when tools are used by adolescents or individuals in crisis without oversight. Moreover, data shared with these apps is often stored by third-party servers, sometimes in jurisdictions without adequate legal protections. As AI systems become more entangled with cloud-based infrastructure and opaque algorithms, patients may unknowingly surrender highly personal data, with little recourse for misuse.

Finally, it is a fallacy to consider AI as neutral. Models are shaped by the data they are trained on and the objectives defined by developers. As shown by Poudel et al. (Poudel et al. 2025), NLP models may fail to detect suicidality in neurodivergent patients or those using non-normative language. In that sense, bias is not a glitch; it is structural. Responsible AI in suicide prevention will require active debiasing, inclusive datasets, and ongoing validation across diverse populations.

Conclusions

Despite the growing body of literature on AI in mental health, its application to suicide prevention remains relatively underdeveloped. Most existing studies focus on predictive algorithms using electronic health records or social media data to estimate suicide risk. While these approaches show promise, they represent only a fraction of the potential contributions AI could make in this field. As this review has outlined, AI could assist in the detection of suicidal ideation during clinical interviews, the training of healthcare professionals, and the support of patients in managing their suicidal thoughts, particularly in settings where access to care is limited.

However, the integration of AI into suicide prevention is fraught with challenges. From ethical concerns about autonomy and data privacy to clinical risks related to automation bias or sycophantic chatbot behaviour, the road ahead requires careful navigation. As we have shown, current use of AI by patients is already widespread—and often unmonitored and unregulated. Ignoring this reality may not only increase the risk of harmful outcomes but also widen the gap between patients and clinicians, especially if outdated paternalistic models of care persist.

It is therefore urgent to move forward in a responsible, evidence-based manner. AI tools intended to support suicidal patients must be subjected to the same rigorous standards as any other clinical intervention: RCTs, regulatory approval by competent authorities (e.g., FDA, EMA), and post-deployment safety monitoring (such as done with pharmacovigilance). In short, the question is no longer whether patients will use AI in mental health—it is whether we, as clinicians and researchers, will shape that use wisely.

Funding Statement

This study did not receive funding from any sources.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work the author(s) used ChatGPT to improve English of the manuscript. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.

Disclosure statement

No potential conflict of interest was reported by the author(s).