The context to large language model and youth mental health

The ubiquitous influence of the internet, particularly among youth, has implications for mental health and cognition that are not yet fully understood (Firth et al., Reference Firth, Torous, Stubbs, Firth, Steiner, Smith, Alvarez‐Jimenez, Gleeson, Vancampfort, Armitage and Sarris2019; Mansfield et al., Reference Mansfield, Ghai, Hakman, Ballou, Vuorre and Przybylski2025). Large language models (LLMs), often deployed as online ‘chatbots’, are deep learning artificial intelligence (AI) platforms trained on textual data to provide conversational style responses akin to interacting with a human. Typically, LLMs centre on a user ‘prompting’, with the application then ‘responding’ (Chatterji et al., Reference Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman2025). The use of LLMs now represent a third generation of AI applications with dynamic human-like fluency and contextual understanding beyond rule-based responses (Hua et al., Reference Hua, Siddals, Ma, Galatzer‐Levy, Xia, Hau, Na, Flathers, Linardon, Ayubcha and Torous2025). The scale of LLM use in a short amount of time is extraordinary, with up to 700 million weekly users of one widely used platform ChatGPT (Chatterji et al., Reference Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman2025). These technologies are increasingly being used to support mental health among young people, thus there is a need to highlight opportunities and inherent challenges of AI based LLMs in this domain.

It is important first to differentiate between LLM chatbots designed for clinical mental health interventions (e.g. Headspace, Youper and Wysa) and powerful, generic, and widely available LLMs (e.g. ChatGPT series and Google’s Gemini) (Torous et al., Reference Torous, Bucci, Bell, Kessing, Faurholt‐Jepsen, Whelan, Carvalho, Keshavan, Linardon and Firth2021). Not surprisingly, mental health contexts are cautiously embracing AI technology in many forms (Torous et al., Reference Torous, Myrick, Rauseo-Ricupero and Firth2020). Out of clinical contexts, AI is also increasingly used for mental health support, most notably among youth who demonstrate a predilection for such technologies on account of their perceived accessibility and anonymity (Mcbain et al., Reference McBain, Bozick, Diliberti, Zhang, Zhang, Burnett, Kofner, Rader, Breslau, Stein, Mehrotra, Pines, Cantor and Yu2025). However, emergent evidence indicates an apparent mental health risk among certain populations when using generic LLMs outside of clinical contexts (Morrin et al., Reference Morrin, Nicholls, Levin, Yiend, Iyengar, DelGuidice, Bhattacharyya, MacCabe, Tognin, Twumasi, Alderson-Day and Pollak2025). It is the coalescence of these phenomenon, an increasing adoption of LLMs among youth to manage mental health, and a new manifestation of risk through third generation LLM platforms which we argue makes for a unique youth mental health phenomena. We therefore set out to contextualise this developing issue, first outlining the context of LLM chatbot style interface use among youth and the simultaneous emergent mental health challenges for this group. We then describe the broad potential for LLM conversational agents and chatbots in a clinical context and the hypothesised risks of generic LLM applications when used to support mental health difficulties.

The relevance of youth mental health and LLMs

There is a notable changing trend in population level mental health among young people, where low to moderate mental health problems have been shown to increase in many contexts over the past decade (Mcgorry et al., Reference McGorry, Gunasiri, Mei, Rice and Gao2025). As an example, the second iteration of the My World Survey showed that rates of low to moderate level depression and anxiety have increased among young people in Ireland between 2012 and 2019 (Dooley et al., Reference Dooley, OConnor, Fitzgerald and O’Reilly2019). The global COVID-19 pandemic has further marked a change in increasing online digital platform engagement, concurrent social isolation, and mental health challenges among youth (Solmi et al., Reference Solmi, Thompson, Cortese, Estrade, Agorastos and Radua2025). The underlying reasons for this youth mental health challenge are multidimensional, with many changing global trends implicated. Online and social media influence is one such trend that has been identified as a key proximal factor driving changes in youth mental health (Mcgorry et al., Reference McGorry, Gunasiri, Mei, Rice and Gao2025).

Indeed, young people are widely acknowledged as adept users of digital technologies such as LLM powered applications. For instance, up to half of all ChatGPT users are reportedly under 26 years of age (Chatterji et al., Reference Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman2025). Moreover, many recent AI innovations and developments are increasingly targeting youth. For example, there is an emergence and increasing influence of embedded AI conversational agents and companions within widely used social media applications targeted at young people (Sun, et al., Reference Sun, Wang and McDaniel2026). Reflecting a propensity for digital solutions, young people may be increasingly turning to online spaces to manage a range of mental health difficulties outside of the clinical context, this includes the use of generic LLM driven platforms and applications (Firth et al., Reference Firth, Torous, López‐Gil, Linardon, Milton, Lambert, Smith, Jarić, Fabian, Vancampfort, Onyeaka, Schuch and Firth2024). For instance, a US based survey of 1,058 young people showed that 13% had used generic LLMs for informal mental health support, with this number increasing to >22% among 18 to 21 year old respondents. Among persons that used such technology for support, more than half reported using it regularly (Mcbain et al., Reference McBain, Bozick, Diliberti, Zhang, Zhang, Burnett, Kofner, Rader, Breslau, Stein, Mehrotra, Pines, Cantor and Yu2025). Additional recent survey data from 10,835 13–17-year-olds in the UK similarly found 25% of the sample had used generic LLMs to support their mental health, showing such technologies are being increasingly relied upon by young people to manage their mental health (The Youth Endowment Fund 2025).

The potential benefit of artificial intelligence and LLMs in mental health

Broadly, AI technologies may be an opportunity for the augmentation of mental health care, including therapeutics (Hua et al., Reference Hua, Siddals, Ma, Galatzer‐Levy, Xia, Hau, Na, Flathers, Linardon, Ayubcha and Torous2025; Torous et al., Reference Torous, Bucci, Bell, Kessing, Faurholt‐Jepsen, Whelan, Carvalho, Keshavan, Linardon and Firth2021, Reference Torous, Linardon, Goldberg, Sun, Bell, Nicholas, Hassan, Hua, Milton and Firth2025), diagnoses (Graham et al., Reference Graham, Depp, Lee, Nebeker, Tu, Kim and Jeste2019), and referral (Habicht et al., Reference Habicht, Viswanathan, Carrington, Hauser, Harper and Rollwage2024). A 2023 narrative review describes AI development as an opportunity to expand capability, improve productivity, and to support the work of practitioners across domains with promise for young people in particular (Balcombe Reference Balcombe2023). An additional review not specific to youth services indicated that specific mental health chatbots can offer a level of acceptability, usefulness, and attractiveness among people that are using mental health services (Abd-Alrazaq et al., Reference Abd-Alrazaq, Alajlani, Ali, Denecke, Bewick and Househ2021). Indeed, ease of access, associated low cost, and inherent anonymity make such platforms an intuitive support tool for young people and adults (Lawrence et al., Reference Lawrence, Schneider, Rubin, Matarić, McDuff and Jones Bell2024). Early literature, predominantly focused on ‘rule-based’ (if/then logic, scripts, decision trees) chatbots demonstrates potential in addressing mental health problems in clinical contexts (Abd-alrazaq et al., Reference Abd-Alrazaq, Alajlani, Abdallah Alalwan, Bewick, Gardner and Househ2019). The ‘digital exhaust’ which is the by‑product of an individual’s interactions with digital systems, can be used by LLMs to profile individuals, and respond accordingly in a therapeutic manner (D’Alfonso Reference D’Alfonso2020). A recent small scale case series intervention showed promising outcomes from a specifically designed rules based platform among youth in Australia with anxiety and depression, where distress and symptoms of anxiety were found to meaningfully reduce following chatbot interaction (Wrightson-Hester et al., Reference Wrightson-Hester, Anderson, Dunstan, McEvoy, Sutton, Myers, Egan, Tai, Johnston-Hollitt, Chen, Gedeon, Moullin and Mansell2025).

Looking to the third generation of AI applications that utilise adaptive systems and LLM driven applications, small sample research shows that an immersive multimodal AI driven therapy leveraging spatial computing, virtual reality along with a digital AI therapy avatar assistance may be both safe and acceptable among adults with mild to moderate anxiety and depression (Spiegel et al., Reference Spiegel, Liran, Clark, Samaan, Khalil, Chernoff, Reddy and Mehra2024). A recent randomised trial of a four-week ‘Therabot’ intervention with personalised capability shows therapeutic potential with reductions in symptoms of major depression, general anxiety, and feeding and eating disorder in over 200 adults with these respective diagnoses when compared to controls (Heinz et al., Reference Heinz, Mackin, Trudeau, Bhattacharya, Wang, Banta, Jewett, Salzhauer, Griffin and Jacobson2025). Powerful generic LLMs may also play a role in mental health support at a population level. As already noted, one UK survey shows that many young people already use LLMs for mental health support (The Youth Endowment Fund 2025). Another global survey of >1,000 users of the AI platform Replika indicated N = 30 case examples where the platform is reported to have prevented a suicide attempt (Maples et al., Reference Maples, Cerit, Vishwanath and Pea2024). Given the newness of this phenomena, robust data from research on youth populations is urgently warranted to determine the extent and value of AI LLMs in supporting mental health both in clinical and non-clinical settings.

Concerns about LLMs and mental health

Outside of clinical contexts, there are emergent reports of adverse mental health phenomena from human interaction with generic LLMs that give rise for concern (Morrin et al., Reference Morrin, Nicholls, Levin, Yiend, Iyengar, DelGuidice, Bhattacharyya, MacCabe, Tognin, Twumasi, Alderson-Day and Pollak2025). The underlying risk appears linked to inherent system-level features of the LLM and the human-level risk profile of the user (Dohnány et al., Reference Dohnány, Kurth-Nelson, Spens, Leuttgau, Reid and Gabriel2026). These phenomena also occur within the context of inherent human attributes such as confirmation bias, motivated reasoning, and anthropomorphism, which may further explain the preponderance of risk among some users (Dohnány et al., Reference Dohnány, Kurth-Nelson, Spens, Leuttgau, Reid and Gabriel2026; Peter, et al., Reference Peter, Riemer and West2025). Reflecting the newness of the phenomena, much of the evidence underpinning recent catastrophic events highlighted in the media are notably emergent or theoretical (Morrin et al., Reference Morrin, Nicholls, Levin, Yiend, Iyengar, DelGuidice, Bhattacharyya, MacCabe, Tognin, Twumasi, Alderson-Day and Pollak2025), and not yet identified as youth specific. Key risk phenomena across these system and human levels are outlined below. It is also relevant to consider that the discussed phenomena below appear to pose a complex, linked, and or bi-directional risk to certain persons.