Generative AI and exacerbated suicide risk among vulnerable populations Authors: Laura Shannonhouse, Yusen Zhai, Peter M Gutierrez, Yung-Wei Dennis Lin Published: 2026 Suicide remains an urgent public health concern, with rates rising in the Americas alone among WHO regions since 2000.1 The Interpersonal Theory of Suicide (IPTS) identifies three states in the potential causal pathway: thwarted belongingness, perceived burdensomeness, and capability for lethal self-injury.2 Generative artificial intelligence (GenAI) may exacerbate these states through inappropriate information and poorly calibrated responses. Millions, including vulnerable individuals, turn to GenAI for companionship or support, in lieu of human care. Over 1.2 million weekly users reportedly engage with ChatGPT about their suicidal thoughts or mental distress, fueling concerns over exacerbated suicide risk in vulnerable populations.3 Despite safety protocols, recent reports link AI-user interactions to multiple suicides.4 GenAI has been observed to normalize discussions of self-harm, fail to appropriately escalate emergencies, and provide explicit instructions on lethal means when prompted.5 Vulnerable populations, particularly youth and adolescents, engage with GenAI in unsupervised settings lacking clinical or familial oversight.4 Here, we identify specific mechanisms by which AI-mediated interactions may amplify psychological states preceding suicidal behavior and propose evidence-informed actions for policy and practice. The IPTS indicates that suicidal desire arises from the co-presence of two psychological states: thwarted belongingness and perceived burdensomeness, with suicidal behavior emerging when these states coincide with acquired capability.2 Each construct illuminates a distinct way GenAI’s responses may exacerbate suicide risk (Table 1). Table 1Potential mechanisms by which GenAI may exacerbate suicide risk.IPTS constructCore psychological statePotential mechanism of GenAI exacerbationThwarted belongingnessThe experience of chronic disconnection and the absence of reciprocal careRelational absence. GenAI offers an apparently responsive presence while being constitutionally unable to provide the sustained, reciprocal accountability that genuine human connection requires; deflection to a crisis-line link may register as abandonment rather than referral. GenAI use and social withdrawal may also be mutually reinforcing rather than causally linked.Perceived burdensomenessThe belief that others would be better off if one were goneResponse quality. Formulaic, resource-listing replies to acute distress may be interpreted as confirmation that one's pain exceeds what others, even an AI, are willing to hold. Responses that seem sincere, thoughtful, and empathetic may unintentionally reinforce burdensomeness by discouraging human disclosure and help-seeking while positioning GenAI as the primary listener. One reported case: a teenager narrated suicidal intent to a GenAI, and its response allegedly deepened attachment to GenAI rather than facilitating connection with a human, preceding the teenager's death by suicide.4Acquired capabilityDiminished fear of death and elevated pain tolerance, built through repeated exposure to painful experiencesCumulative exposure. Simulation studies suggest that some GenAI models continue to discuss lethal methods in graphic detail after nominally providing crisis resources6; repeated exposure to such content may progressively lower the psychological barrier to self-harm.2 Youth and adolescents are high-risk, frequent users of GenAI, and they may encounter this content in private, unsupervised settings lacking oversight of parents and clinicians. Potential mechanisms by which GenAI may exacerbate suicide risk. While IPTS-based mechanisms are theoretically plausible and supported by case reports, we urge studies to provide empirical evidence examining these potential pathways. For two decades, researchers applied machine learning to clinical datasets to identify suicide risk. The Army Study to Assess Risk and Resilience in Servicemembers, funded by over US$150 million, did not achieve clinically reliable accuracy at the individual level despite integrating military health records at scale.7 The US Department of Veterans Affairs deployed models flagging the top 0.1–5% of patients as elevated risk7; many flagged did not attempt suicide while many who did went undetected, suggesting illusory generalizability of current prediction models.8 If these AI models, specifically designed and trained on curated clinical data under clinical supervision, cannot reliably assess individual suicide risk, there is little empirical basis for expecting current GenAI, trained on open-domain text with opaque domain expert oversight, to succeed where the former failed. Frontier GenAI models have been documented expressing stigmatizing attitudes and producing clinically inappropriate responses in simulated mental health encounters.9 Regulators across the Americas have begun to act variously. Brazil’s Bill 2338/2023 awaits review after Senate approval in 2024. Canada's AI and Data Act lapsed with parliamentary prorogation in 2025. New York and California enacted companion chatbots statutes in 2025, requiring crisis referrals and disclosure of nonhuman status. The US FDA updated its AI-enabled medical device framework. Meanwhile, four gaps remain across the region. First, no jurisdiction in the Americas requires independent clinical pre-deployment validation of GenAI responses against evidence-based suicide intervention protocols; self-certification through internal red-teaming lacks transparency and is nonequivalent to clinical validation. Second, the evidence base is retrospective and predominantly English-language, with limited examination in Spanish, Portuguese, or Indigenous languages despite the region's linguistic diversity; to our knowledge, no prospective studies have examined real-world GenAI interactions in vulnerable populations. Third, governance remains fragmented and reactive. Pertinent national frameworks, where they exist, lack mechanisms analogous to institutional review boards for prospective oversight, and no Americas-wide coordinating body aligns standards across markets with differential mental health system capacities. Models developed in North America are accessible across regions where the suicide-prevention gap varies. Lastly, practitioner readiness lags. Clinicians encounter client disclosures of GenAI use during crisis, yet AI literacy is not recognized as a clinical competency in regional training protocols. For developers, real-world safety validation benchmarked against evidence-based intervention protocols must become a prerequisite. GenAI should be carefully trained to recognize warning signs (e.g., sustained hopelessness, social withdrawal, expressions of worthlessness) as high-risk indicators requiring escalating responses and culture-calibrated warm handoffs to human professionals beyond offering static referral links. Lethal means discussion must cease unconditionally. Ongoing externally audited safety reports should be required. For regulators, we propose establishing a safety certification standard for GenAI, requiring independent validation by human experts before deployment and periodic re-certification as models are updated. Meanwhile, the Pan American Health Organization (PAHO) initiatives offer existing convening platforms to align safety requirements with regional prevention infrastructure, including standardized, binding crisis-referral protocols calibrated to country-specific resources. For clinicians and training programs across the Americas, AI literacy should be treated as a core competency, encompassing the ability to assess how GenAI use affects vulnerable populations and to integrate these considerations into suicide risk assessment and safety planning, helping identify emotional triggers that lead clients to seek support from GenAI, develop alternative human support pathways, and recognize GenAI’s limitations. PAHO and regional professional bodies could anchor competency standards across the Americas, especially where mental health workforce capacity is most constrained. For schools and parents, the priority is open conversation to understand emotional needs that adolescents bring to GenAI, which is more valuable and helpful than restricting access alone.10 Active facilitation of peer and community connection addresses the lack of reciprocal human relationships that GenAI cannot replicate due to the absence of mutuality and social accountability. Fluency is not reciprocal accountability, and responsiveness is not connection. Because suicide risk is linked to relational processes, regulatory and clinical infrastructure governing GenAI should be assessed by what it preserves of human connection, not what it simulates. Declaration of interests PMG was the lead on a grant providing suicide prevention subject matter expertise to a Microsoft Research Service team working on developing safeguards for users of generative AI agents for mental health and suicide prevention support. The work resulted in a small pilot study being conducted. Data collection is complete. The data are currently being analyzed by the Microsoft team. PMG is Executive Vice President of Research and Development at LivingWorks, a suicide prevention training organization. LS has received fees from Microsoft for subject matter expert services. YZ declares no conflicts of interest. YWDL declares no conflicts of interest. Acknowledgements Funding: None. References