How effective is artificial intelligence (AI) in youth mental health interventions?
In This Section

How effective is artificial intelligence (AI) in youth mental health interventions?

Last updated: April 2026 (includes research published from 1980 to April 2026)

This living systematic review summarises research evidence on interventions, primarily chatbots, that use artificial intelligence (AI) to support the treatment of mental health issues in young people. It also reports on insights from young people, families, carers and supporters, and youth mental health practitioners. The review is written for professionals who support the mental health of young people in clinical settings. It is designed to support shared decision-making with young people and their families, carers and supporters, as part of responsive and person-centred care. It may also aid decision-making for policymakers, funding bodies, government agencies and AI-program developers. As a living review, it will be periodically updated to incorporate emerging research and developments in this rapidly evolving field.

What is this review about?

This review focuses on interventions that use AI for the treatment of mental health challenges in young people. AI uses advanced computational programming to mimic human-like cognitive processes and tasks. These include learning, reasoning, problem solving, drawing inferences and generalisation.(1, 2)

With increased levels of mental ill-health among young people and increasing pressure on the mental health system, existing services aren’t able to meet the needs of all young people.(3) AI is being adopted in many areas of Australian health care, including mental health, and is used in domains such as diagnosis and treatment, and in administration processes like note taking.(4) Young people are often the first to adopt new digital technologies, and many have already embraced AI tools.(5) The use of AI as part of youth mental health treatment has the potential to overcome existing barriers to care, however limited evidence exists about whether it is safe and effective for improving mental health outcomes for young people. This living review aimed to look at the evidence for the safety and effectiveness of AI-based interventions used to treat mental health challenges in young people.

Although we were open to reviewing a range of AI-based interventions, we only identified controlled trials evaluating AI-based conversational agents (chatbots). No other types of AI-based interventions were found to have had controlled trial evidence published at the time of this review. For the purpose of this review, the term “chatbots” refers to “AI-based chatbots”.

For more general information about AI and youth mental health see the Orygen fact sheet here.

review snapshot

This living review provides a summary of evidence published up to 30 April, 2026. Twelve controlled trials were identified from comprehensive database searches and systematic screening methods (6-17). The 12 trials tested the use of 13 chatbots with 2620 young people with elevated symptoms of depression, anxiety or loneliness.

Overall, the research shows that chatbots used in the treatment of mental health issues in young people:

  • may reduce depression and anxiety symptoms. However this finding is based on low certainty evidence, and it's unclear whether the improvements are large enough to matter to young people.
  • have an unclear safety profile at this stage. While five studies reported no adverse events attributable to AI-tools, safety was not measured or reported in the other seven studies.

As a result, it is not yet possible to be confident that chatbots are safe and effective for the treatment of mental health challenges in young people. As more research is conducted, confidence is likely to increase and a firmer conclusion may be reached.

When we spoke to young people and families, supporters and carers, they told us that:

  • AI tools may improve accessibility to youth mental health support by reducing barriers to care, and may provide useful in-the-moment or between-session support. However, AI was consistently viewed as a supplementary tool rather than a replacement for professional care or human connection.
  • there are concerns regarding privacy, data governance, crisis response, accuracy and cultural sensitivity, which contribute to hesitation toward AI-supported mental health care.  

More detail about the included research studies, the perspectives of young people and carers, and implications for practice and policy can be found in the sections below. For more information on how we conducted this review and a table summarising key information about the included studies, see the ‘Results in Detail’ section.

AI-based chatbots work by analysing or interpreting what a person says – for example recognising topics, emotions, or signs of distress – then providing an automated conversational response.

Some chatbots respond by selecting a reply from a library of pre-written responses created by clinicians or researchers (pre-scripted chatbot). Other chatbots create new responses in real time using patterns learned from large amounts of human language (generative chatbot). A third type combines approved therapeutic content with AI-generated conversational wording (hybrid chatbot).

It important to understand these different chatbot types because each may involve different benefits and risks. Pre-scripted chatbots provide predictable and clinically consistent responses and may be easier to monitor for safety, but they can feel less conversational and may struggle when a young person says something unexpected or outside the chatbot's programmed content.

Generative chatbots may provide more natural, flexible and personalised conversations compared to pre-scripted chatbots. They are trained on large amounts of human language, and may draw on a broader range of both conversational and potentially clinically relevant information. However, they have potential to produce inaccurate, misleading, or inappropriate responses.(18)

Hybrid chatbots aim to combine the safety of approved therapeutic content (pre-scripted information, or a known source like a CBT manual or framework) with the flexibility of conversational AI, but may still carry some risks associated with generated language.

The 12 included studies evaluated 13 chatbots.

  • Seven were pre-scripted chatbots, one was a fully generative chatbot and five were hybrid chatbots.
  • All 13 chatbots were used by young people as standalone interventions, meaning the chatbot was used as the only intervention and not part of standard mental health treatment.
  • Young people used the chatbots independently, without direct real-time clinician or researcher involvement during chatbot conversations. Consistent with the standalone design, no additional therapeutic support was offered between chatbot interactions.
  • Across studies, young people used the chatbots for periods ranging from one to 12 weeks. The expected frequency of use also varied, with some studies asking young people to use the chatbot daily, others weekly, and some allowing unlimited or self-directed access.
  • Each chatbot was evaluated to determine whether it could improve symptoms of mental ill-health. Of the 13 chatbots, 11 were classified as mental health-focused chatbots and two were classified as general-purpose chatbots.

Nine studies included high school or university students and one included young adults from the community. Each of these studies recruited young people with elevated mental health symptoms (e.g., young people had to have at least mild symptoms of depression, anxiety, psychological distress or loneliness). Many studies excluded young people with a mental health diagnosis, higher severity symptoms or those receiving mental health treatment. Only two studies specifically recruited adolescents from clinical populations. In one of these studies adolescents with a new diagnosis of anxiety or depression were recruited, and in the other study, adolescents who were triaged to outpatient care for depression or anxiety were recruited.

This suggests that most of the available evidence has been conducted in young people with elevated symptoms, with less research in clinical or higher severity populations.

Overall, the average age of young people across studies ranged from 14 to 23 years. On average there were more female participants than male participants. No studies reported on non-binary or trans status of the young people they recruited. While ethnicity was reported in some studies, information on First Nations status was not reported.

Of the 12 included studies, only one study was conducted in Australia. Most were conducted in China (six studies), two were conducted in the USA, and one each in Indonesia, Israel and Ukraine. More research is needed in Australian settings to explore feasibility and cultural appropriateness. Interventions were often delivered through secondary schools and universities, with only two studies being conducted in clinical settings (a children's hospital and paediatric care clinics).

Each study aimed to see if the chatbot improved mental health symptoms as a primary target, measuring either depression symptoms, anxiety symptoms or feelings of loneliness. These symptoms were measured on a range of self-report scales. The most common scales were the Patient Health Questionnaire-9 (PHQ-9) for measuring depression, and the Generalised Anxiety Disorder-7 (GAD-7) scale for measuring anxiety. These scales tap into various symptoms associated with depression and anxiety and can show how severe these symptoms might be. For example, the PHQ-9 gives a score between 0 and 27, with higher scores suggesting higher levels of depression.

Safety can refer to (1) the effects the chatbot has on a young person (participant safety), (2) whether the chatbot itself provides safe, appropriate and reliable responses (chatbot response safety), and (3) whether the chatbot has an effective in-built risk identification and safety pathway when a young person conveys clinical risk. Current evidence is insufficient to determine whether chatbots used in youth mental health treatment are safe or unsafe across these domains. As more studies are conducted and published, more safety data may become available, and a firmer conclusion may be possible.

Participant safety: In clinical trials, this is measured as an adverse event and can include worsening psychological distress, increased suicide risk, or hospitalisation. Only five of the 12 studies monitored and reported participant safety outcomes. The remaining seven either did not measure or report these outcomes. Of the five studies that reported safety outcomes, three found no adverse events occurred during the intervention period. Two reported adverse events such as increased emotional distress, suicidal ideation, or emergency department presentations. However, these events occurred at similar rates in both the intervention and control groups, suggesting they were unlikely to be caused by the AI tool itself.

Chatbot response safety: Examples of potentially unsafe responses include “hallucinated” or false responses, unsafe advice or misleading information, or inappropriate language or content. For this review, we were open to any form of chatbot response safety monitoring and reporting; however, none of the 12 studies formally evaluated these outcomes.

In-built risk identification and crisis responding: This refers to situations where a chatbot detects signs of risk or crisis in what a young person types and automatically takes an action, such as referring them to a clinician, mental health service, or crisis helpline. Two chatbots reported having an in-built risk identification and crisis response pathway. However, these processes were often not described in enough detail to allow verification or evaluation of how accurately or effectively they worked.

A further note on safety: Chatbot response safety can also include cultural and identity safety. This refers to whether chatbots respond safely and appropriately to young people from diverse cultural backgrounds, identities, and lived experiences, without reinforcing stereotypes, bias, or discrimination. We considered both direct evaluations of chatbot responses for this type of safety, and whether chatbots were explicitly trained on human language reflecting diverse cultural and identity backgrounds. Across the 12 trials, we did not find evidence that cultural safety had been formally evaluated.

Do chatbots reduce mental health symptoms overall?

Overall, chatbots appeared to improve both depression and anxiety symptoms, however it is unclear if these improvements were large enough to matter to young people.

The current evidence was rated as low certainty. This means that future studies may change the conclusions of this review. In making this judgement, we considered several factors relating to the quality, consistency and precision of the available evidence.

We also examined whether results differed depending on what the chatbots were compared against. Compared to no treatment or wait-list conditions, chatbots appeared to improve symptoms of depression and anxiety. However, when compared to attention control conditions (interventions designed to involve a similar amount of time, attention, or engagement from a young person, but without the active therapeutic components) it was unclear whether chatbots improved depression or anxiety symptoms.

Compared to alternative treatments, such as human-delivered Cognitive Behavioural Therapy (CBT), chatbots appeared to produce similar reductions in depression and anxiety symptoms. However, these comparisons were based on only a small number of studies (four for depression and two for anxiety), meaning the findings remain uncertain.
 

Do chatbots work in clinical and non-clinical populations?

Most studies were conducted in young people with elevated symptoms of mental ill-health outside of a clinical setting. There were only two studies in clinical populations such as those with a diagnosis or help-seeking for mental ill-health.

Chatbots appear to improve anxiety and depression symptoms for those with elevated depression and anxiety symptoms (at least mild symptoms). For clinical populations, it was unclear if chatbots improved symptoms.
 

Does the type of chatbot matter?

For pre-scripted chatbots (seven), it was unclear whether they improved depression and anxiety symptoms. Similarly, for the single generative chatbot, it was unclear whether it improved symptoms. Hybrid chatbots (five) appeared to produce an improvement in depression and anxiety symptoms. 

When compared against one another, there appeared to be no clear benefit of one chatbot type over another, however there were likely too few studies in these categories for the comparison to be meaningful.
 

As part of this project, an environmental scan was conducted to understand how young people are using AI-enabled apps for their mental health in real life, and which apps are readily accessible to them. When compared to the findings of this living review, a clear gap emerges between the AI tools being studied in research and those that young people are actually using. Specifically:

•    Most of the AI-based interventions evaluated in the research were pre-scripted chatbots. However, consultations with young people indicated that the most commonly used AI tool for mental health support is ChatGPT, a generative model that produces novel responses.

•    Only one of the 13 tools in the review was fully generative. Fully generative tools like ChatGPT have more open-ended and variable responses than pre-scripted chatbots. These may have a higher risk of producing an inaccurate, unhelpful or potentially harmful responses than a tool that uses more predictable pre-scripted responses, due to having fewer built-in constraints.(18) However, fully generative tools may also provide a more natural and engaging user experience.

•    In all included studies, the chatbots were delivered as standalone interventions, as opposed to one part of a multicomponent intervention. 

•    The rapid pace of AI development and availability is outstripping the evidence base. By the time many studies are published, the tools or models evaluated may already be outdated, superseded or no longer widely used.
 

Young people’s reflections on the use of ai in youth mental health treatment

We spoke to 11 members of Orygen’s National Youth Advisory Council (YAC) – young people aged 18-25 years – to understand their perspectives on AI tools for youth mental health. From these discussions, it was clear that young people were most likely to turn to ChatGPT specifically when seeking mental health support from AI tools. They were unlikely to use mental health-specific AI tools, unless directed to them by a mental health professional, which would make them feel the tools were safe to use and endorsed by a professional. 

Young people indicated they were most likely to use AI for general advice and information about mental health issues, as opposed to in times of crisis. Of the six YAC members who identified as having lived or living experience of mental ill-health, three reported that they would use, or have used, AI for their mental health before seeking professional treatment, care or support. Furthermore, two people said they would use, or have used, AI instead of professional treatment, care or support.

The group highlighted many potential benefits of AI, including reducing accessibility and societal barriers (such cost, location and long wait times) to accessing mental health care. They also highlighted that AI could reduce stigma and shame - giving young people a sense of “anonymity” and access mental health support in their own time and space. AI also provides a space for young people to “offload” without feeling like they are burdening friends or family.

However, there was an overall sense of hesitation to use AI for mental health support. The six young people interviewed were concerned with data safety, privacy, and regulation. They felt little is known about where data is stored, who has access to it, and who is responsible if there were to be a data breach. The six people were also concerned that AI is not adequately equipped to respond during emergency and crisis situations effectively. With recent suicide deaths reported in the media, they highlighted that AI does not have sufficient crisis referral pathways and does not respond in a way that a mental health professional would. Lastly, they were concerned with the type of advice provided by AI, highlighting that it is sometimes overly agreeable, inaccurate, and potentially not culturally safe or sensitive, creating a sense of untrustworthiness. 

“As an international student, I have used ChatGPT many times because it is hard to afford a psychologist. It can learn to be so kind, and you can ask about things from another perspective. However there is always the ‘push-pull’ of thinking about the safety aspect and if the data ever gets leaked.”
— Young person

young people’s reflections on the findings of this review

We also spoke to young people about the research findings from the living review. 

From their perspective, the lack of monitoring and reporting of adverse outcomes in the majority of studies stood out. Young people further emphasised that even with generative AI, effective and safe use depends on the individual, including their ability to recognise limitations and make informed decisions about when AI is appropriate. For example, asking for strategies to manage a stressful situation, but not asking for advice about medication.

The young people largely reported that the results made them reflect on their own experiences of using AI for mental health support. For example, they described AI being helpful for providing in-the-moment support, especially when other supports are unavailable. However in their view, AI does not teach structured psychological skills or provide the depth of support needed for long-term improvement. Therefore the young people agreed that AI cannot replace professional treatment, care and support but was instead seen as a supplementary tool, helpful for initial support or between sessions. This aligns with the National Mental Health Consumer Alliance position paper, which asserts that AI tools should not replace or be implemented as a substitute for locally available relationship-based care.(19) The observations from young people about viewing AI as helpful only in the short term reflected the research finding that AI reduced depression, anxiety and loneliness immediately following the interventions, but more research is needed on whether the effects are maintained long-term.

“AI is helpful in a short-term situation but not for long-term treatment. When I use it, it doesn’t teach me the proper tools that a psychologist would teach you. It doesn’t replace a psychologist.” — Young person

The look and feel of some AI tools used in the research studies was consistent with chat-based interfaces young people were already familiar with, such as WhatsApp and WeChat, which they reported as increasing perceived trustworthiness and confidence to use the apps.

What young people want clinicians to know

Young people identified several key considerations for clinicians. They highlighted that many young people are already using and finding benefit from using AI for their mental health, and that clinicians should acknowledge this without judgement. However, there is an important role for clinicians to support safe use of AI. This includes building the digital literacy of young people about the limitations of AI, and explaining when using AI is appropriate and when professional care is recommended. Young people highlighted that some might be hesitant to use AI due to its environmental and economic/poverty impact.

Young people raised opportunities for AI to be integrated into care, such as having the ability to use AI in between sessions with their mental health professional, with the option for the clinician to review these interactions to reduce the burden of re-telling distressing experiences.

They emphasised the importance of clinicians being confident in their own understanding of the AI tools available, including data privacy considerations. Trust in AI tools among young people may be strengthened through clinician involvement in their development. Clinicians can then play a key role in advocating for the tools’ features and benefits, as well as supporting understanding of potential risks and limitations, enabling young people to engage with them safely. Alongside clinician engagement, meaningful involvement of people with lived experience of mental health challenges should be at every stage of digital mental health tools’ governance and development, to further strengthen trust and ensure these technologies meet the needs of those they are designed to support. This notion was strongly asserted in the recent position paper from the National Mental Health Consumer Alliance.(19)

family, carer and supporter reflections 

We spoke with five Family Peer Workers from headspace centres located in Melbourne’s north and north-western suburbs to understand their perspectives as people with lived and living experience of supporting a young person with mental health challenges, and as lived experience professionals, on the use of AI in youth mental health treatment.

Family peer workers were largely unfamiliar with AI in youth mental health treatment and expressed hesitancy, noting that many families are trying to reduce young people’s use of digital tools/devices. Their reactions centred on concern, fear, and sadness about young people seeking validation or advice from AI instead of humans, particularly given risks for those with complex issues such as psychosis. They also questioned whether AI could ever replicate the comfort, understanding, and relational value provided by lived experience workers.

When asked about their concerns related to AI-assisted apps and tools, families said they would feel uneasy about an AI-assisted therapy app unless they had clear information regarding its safety, data use, personalisation, and how risks would be managed, especially in complex or high-risk situations. Their concerns centred on inaccurate or generalised advice, lack of human nuance, cultural bias, confidentiality and data misuse, unclear accountability, and the challenge of contacting an “app” when worried about a young person. While some could see potential benefits, they emphasised the need for strong safeguards, cultural responsiveness, and clear pathways for family involvement in decision-making.

Conversely, families recognised that AI-assisted tools are attractive from a financial/resource perspective and could increase accessibility and scalability, offering timely support that helps clarify, validate, and normalise young people’s experiences. Some also saw value in these tools to support families to do therapeutic work at home, and saw the benefits that a “gamified” care experience could provide to keep young people engaged.

“As family peer workers we always talk about how most of the work is done at home. AI could play a role helping families to support their young person at home by reminding them of psychoeducation, supporting them to build skills and for between sessions seeing their family peer worker.”
— Family peer worker

Families said AI tools would only feel acceptable if they operated under strong human oversight, with clear clinical involvement and were used strictly as an adjunct, not a replacement, for mental health care. Trust, reliability, and strong engagement features (such as gamification) were seen as essential. Use cases such as tools that meaningfully support communication and provide practical, in-the-moment strategies for common issues like self-harm were suggested as valuable or acceptable examples of interventions with an AI component. The family peer workers emphasised that AI could be helpful if it enhances existing care and strengthens family-young person support rather than replacing human connection.

In terms of how the peer workers were using AI-assisted apps and tools to help support their own mental health and wellbeing, they described using AI mainly as a starting point for researching strategies, normalising experiences and accessing basic psychoeducation – particularly during late-night moments when other supports weren’t available. However, many were reluctant to rely on it, emphasising the irreplaceable value of human experience and noting that AI feels untrustworthy in crises. They felt anonymity and convenience were benefits, but scepticism, safety concerns, and uncertainty about accuracy remained key barriers.

Families emphasised that youth mental health workers need to understand AI’s limitations, use clinical discretion, and clearly communicate to young people and families where information comes from and where data goes if they are using AI in treatment. They stressed the importance of genuine informed consent, not assuming young people fully grasp the technology, and ensuring safeguards and disclaimers are always in place. Above all, they highlighted that lived human experience, relational understanding, and transparent clinical reasoning remain essential and cannot be replaced by AI.

what do these findings mean for youth mental health practitioners?

We spoke to a group of eight youth mental health professionals, including clinicians currently working with young people, to understand their perspectives on AI tools for youth mental health. Professionals seemed to understand that young people like to use AI because it is easily accessible, low-cost, always validating and “on your side”, with less risk of the stigma and shame that some young people feel when accessing mental health services. It was also a common perception that young people generally know more about AI than clinicians do, suggesting a need for upskilling and educating the youth mental health workforce.

Many agreed that AI tools specifically designed as mental health supports could be useful as an adjunct, or be complementary to in-person treatment and therapeutic work. However, professionals had significant concerns about lack of clinical oversight, the accuracy of advice AI produces, and whether responses were customised for Australian culture. They also had concerns about governance, what happens when a young person discloses risk (e.g., mandatory reporting legislation), data privacy and young people’s understanding of the limitations/boundaries of their relationship with AI tools. These views were similar to those reported in the 2025 Safe and Responsible Artificial Intelligence in Health Care Legislation and Regulation Review)(20), which involved public consultation. The review concluded that whilst existing legislation can largely accommodate AI technologies in health care, minor amendments and economy-wide guardrails are needed to strengthen safety. It also emphasised the importance of an evidence-based approach to AI deployment to realise productivity benefits equitably whilst managing risks in high-risk settings, such as mental health care.

The findings from this review suggest some chatbots that are specifically designed as mental health support tools may be beneficial for young people who have mild symptoms of depression and/or anxiety. However, the strength of the evidence is not at a level where we can confidently say there are benefits, and due to the lack of data about adverse events/safety, it is presently recommended that clinicians use caution when considering AI-based chatbots for use in treatment with young people.

“A chatbot won’t ever be able to do what a therapist can do.”
— Youth mental health clinician

Tips for clinicians working with young people using AI TOOLS

While there are many unknowns about the impact of AI tools on young people’s mental health, discouraging use of AI outright is not recommended – blanket discouragement may reduce young people’s willingness to share their experiences of using AI tools. Young people may also be using AI tools appropriately, and for things unrelated to mental health.

To support safest possible AI use, clinicians working with young people should:

  • Take an open, curious, non-judgemental approach, focus on rapport and understanding why and how the young person uses AI.
  • Provide psychoeducation on safe and critical use, including how AI generates responses, potential bias based on user input and limitations of AI (not a replacement for therapy).
  • Encourage young people’s critical thinking, for example asking themselves “where is this information coming from and how has it been generated?” “who benefits from this tool?” and “how does it make me feel after using it?”
  • Support young people to reflect on their usage patterns (e.g., time, reliance, does it impact their behaviour) and notice if it keeps them engaged or in a repetitive loop.
  • Highlight suitable vs less suitable use of AI, e.g., it can be helpful for generating or honing ideas, reflection, learning skills, but is less appropriate for self-diagnosis and as a primary source of support for mental-ill health.
  • Encourage balance with real-world supports by reinforcing existing coping strategies and promoting human connection.
  • Where possible, explore AI use together in session - experiential learning can be more effective than explanation.
  • Help young people improve how they use AI (e.g., asking better questions, requesting evidence-based responses).

“AI especially in YMH, it is easy to think of connotations such as unhelpful, dangerous, but it could value add to our role as clinicians and it has a role in communities where people don’t have access, or the money to access regular sessions.” — Youth mental health clinician

working with first nations young people

The review did not identify any AI-enhanced mental health tools designed specifically for Aboriginal and Torres Strait Islander young people. However, emerging Australian research points to both the potential and the risks of using AI with First Nations populations. AI tools could offer benefits such as around-the-clock access to support and reduced barriers to help-seeking for First Nations communities. At the same time, caution should be used as tools trained on overseas data may not reflect First Nations understandings of mental health and wellbeing, and risk perpetuating existing biases if not developed with communities.(21-23)

There is strong agreement across the sector that any AI tools developed for First Nations communities must be co-designed with and governed by those communities, grounded in social and emotional wellbeing frameworks, and protective of Indigenous data sovereignty. This represents an important opportunity for clinicians and policymakers to advocate for tools that reflect First Nations voices and knowledges from the outset.

Key resources for working with Aboriginal and Torres Strait Islander young people include:

  • 13YARN is a national service providing telephone access to Aboriginal or Torres Strait Islander Crisis Supporters 24 hours/7 days.
  • WellMob provides social, emotional and cultural wellbeing online resources for Aboriginal and Torres Strait Islander People, including workforces.
  • AIMhi-Y is a digital wellbeing app for First Nations young people aged 12–25, co-designed with young people and Elders in the Northern Territory. While it doesn’t use AI, it embeds Elder guidance and connection to Country into a brief, self-guided intervention.
  • Social and Emotional Wellbeing (SEWB) framework provides an evidence-based model for understanding Aboriginal and Torres Strait Islander mental health and wellbeing, including specific domains related to connection to land, culture, ancestry, family, and community.

Note: This review focused specifically on controlled trials, a research method which is rooted in Western scientific paradigms which may not align with First Nations ways of knowing, doing, and being. (24,25)  As such, it is difficult to determine from this evidence base whether any AI-enhanced tools have demonstrated effectiveness specifically for First Nations young people.

where to from here?

Research

  • There is a clear need for more rigorous and independent evaluation of generative AI tools , especially those with fewer constraints such as ChatGPT.
  • Research in this area should consistently report on safety, including adverse events or unintended harms associated with AI tool use.
  • Current research is limited to mainly non-clinical samples and a small number of conditions. Broader populations, diagnoses and longer-term outcomes are needed.
  • Research within priority populations including First Nations young people, LGBTQIA+ young people, young people from culturally and linguistically diverse backgrounds and from rural and remote areas is required.  
  • Importantly, given all tools were tested as standalone interventions, there is an opportunity for research to examine how AI can function as an adjunct to clinical care. This could include research exploring implementation within real-world service delivery models, with a focus on acceptability, feasibility and process outcomes in addition to efficacy.

Policy

  • In the absence of strong evidence, clinicians need practical guidance on how to respond to and safely integrate AI use in care where appropriate.
  • Funding and infrastructure are needed to support rapid, ongoing evaluation of AI tools, particularly those already in use by young people.
  • There is a need for accessible AI literacy resources with information about how AI works in the tools clinicians recommend, its benefits and limitations and how to identify safe tools. These could be publicly available and shared by clinicians with young people who are using AI for mental health support.

App development

  • Consultation with priority groups and communities during app development and testing is essential to ensure that AI tools are meaningful to users and do not further perpetuate existing mental health inequity.
  • The requirements for high-quality controlled trials do not align well with the rapid pace of AI innovation. Partnerships between developers and researchers could support ongoing, real-world evaluation of publicly available tools. This could allow the tools to be extended beyond their intended use cases (i.e., mental health treatment) in controlled environments, generating evidence post-release. This approach could help keep evaluation responsive to innovation while maintaining oversight of safety, effectiveness, and real-world use, and may support future eligibility for regulation by the ARTG.

take-home message

At present, there is limited evidence to suggest that some chatbot interventions, designed as mental health support tools, may be beneficial for young people who have mild symptoms of depression and/or anxiety. However, due to the small number of controlled trials, the variability in study methodology, and lack of data on adverse effects, it is not possible to say for certain these tools are helpful, and we cannot rule out potential harms. It is important to note that all studies except for one were conducted overseas, calling into question the generalisability of findings to Australian contexts. This literature review also focused only on AI-tools explicitly targeting mental health issues in young people and did not examine research related to non-treatment-based AI-tools.

Common concerns raised by young people, families and youth mental health practitioners which were not addressed by the literature centred around data privacy, safeguards, accuracy of information and lack of regulation. Our environmental scan highlighted the significant disconnect between AI treatment-based tools with published scientific evidence and AI-tools being used for mental health support in the real-world. The challenge within this field is reconciling competing demands: balancing the race to develop and market ever-evolving cutting-edge AI-tools and thorough testing of such tools via sound research methods.

It is inevitable that some young people will continue to use AI for mental health support, and that generative as opposed to pre-scripted AI-based tools are the way of the future. Governments globally are rallying experts to understand and address the broader issues of AI, including AI in health, and we are aware of funding bodies making this a priority research area. We expect that researchers, AI-developers and regulatory bodies will continue to work together to ensure AI-tools are safe and helpful for young people with mental health issues. However, until more evidence emerges, we currently recommend health professionals use caution when considering AI-treatment based tools in the care of young people with mental ill-health.

For more information about AI, youth depression and anxiety:

Orygen fact sheet: artificial intelligence in youth mental health fact sheet – information about artificial intelligence and their role in youth mental health support.

Orygen environmental scan: artificial intelligence and young people’s mental health – an environmental scan of available apps and real-world use

Orygen depression resources – designed for youth mental health professionals, includes clinical practice guides and manuals, online learning modules, evidence summaries and more. 

Orygen anxiety resources  – designed for youth mental health professionals, includes clinical practice guides and manuals, online learning modules, evidence summaries and more.

Artificial intelligence (AI): AI uses advanced computational programming to perform tasks that normally require human-like processing such as learning, reasoning, problem solving, drawing inferences and generalisation. For this review, we specifically defined AI as requiring machine learning or other statistical methods to learn patterns from large amounts of data.

Chatbot: a computer program designed to simulate conversation through text or voice interactions, sometimes called a conversational agent. (Chatbots that were not AI-based were not reviewed as part of this project).

AI-based chatbot: A chatbot that uses machine learning-based methods either to understand what a person says and/or in generating conversational responses.

  • Pre-scripted chatbots respond using pre-written clinician- or researcher-approved responses selected from a library of prepared replies. Although some systems may use AI to interpret what a young person types, the responses themselves are fixed and pre-written rather than newly generated.
  • Generative chatbots create novel responses in real time based on patterns learned from large amounts of human language. Rather than selecting from a fixed library of responses, these systems generate new conversational replies during the interaction.
  • Hybrid chatbots combine approved therapeutic content with AI-generated conversational language. In these systems, the chatbot’s responses may be guided or constrained by clinician-approved material, therapeutic frameworks (such as CBT), pre-written content, retrieval of approved information sources (sometimes called retrieval-augmented generation or “RAG”), or prompting instructions that shape how a large language model responds (prompt engineering). The chatbot then uses generative AI to present this information in a more conversational or personalised way.

Control group: a comparison group used for testing whether a program of interest is associated with a significant improvement in mental health outcomes. Control groups in this review could be no intervention, an attention control or another mental health treatment that didn’t use AI.

Controlled trial: a research study that includes a control group as well as an intervention group who are assigned to the AI-based treatment.

See more here.

This review involved a systematic search of academic literature and used knowledge translation principles to consider evidence from lived experience and practice wisdom alongside the research evidence. Data from controlled trials was found using the Evidence Finder and by searching databases (EMBASE, PsycINFO, Medline, Cochrane CENTRAL, ACM-Digital Library) for research published from 1980 to the current date of this review. Included studies focus on treating mental health issues in young people aged 12 – 25 years. Studies were only included if the intervention involved AI-tools used as part of a young person’s treatment.

1.             Thakkar A, Gupta A, De Sousa A. Artificial intelligence in positive mental health: a narrative review. Front Digit Health. 2024;6:1280235.

2.             Kaplan J. Artificial intelligence: What everyone needs to know.: Oxford University Press; 2016.

3.             McGorry PD, Mei C, Dalal N, Alvarez-Jimenez M, Blakemore S-J, Browne V, et al. The Lancet Psychiatry Commission on youth mental health. Lancet Psychiatry. 2024;11(9):731-74.

4.             Dehbozorgi R, Zangeneh S, Khooshab E, Nia DH, Hanif HR, Samian P, et al. The application of artificial intelligence in the field of mental health: a systematic review. BMC psychiatry. 2025;25(1):132.

5.             Firth J, Torous J, Stubbs B, Firth JA, Steiner GZ, Smith L, et al. The “online brain”: how the Internet may be changing our cognition. World Psychiatry. 2019;18(2):119-29.

6.             Gleason MM, Flom M, Rapoport S, Williams A, Birch A, Wells NK, et al. A relational agent intervention for adolescents seeking mental health treatment: Outcomes from a randomized controlled trial within a children’s outpatient hospital. JAACAP open. 2025.

7.             He Y, Yang L, Zhu X, Wu B, Zhang S, Qian C, et al. Mental health chatbot for young adults with depressive symptoms during the COVID-19 pandemic: single-blind, three-arm randomized controlled trial. Journal of medical Internet research. 2022;24(11):e40719.

8.             Indrayanti I, Salsabila AK, Amrita V, Alhaddad MM, Saskia AB, Ramadhani DP. PsyBot: A Randomized Controlled Trial of WhatsApp-Based Psychological First Aid to Reduce Loneliness Among 18–22-Year-Old Students in Yogyakarta, Indonesia. SSM-Mental Health. 2025:100504.

9.             Jiang J, Yang Y. GymBuddy and Elomia, AI-integrated applications, effects on the mental health of the students with psychological disorders. BMC psychology. 2025;13(1):350.

10.           Johnson C, Egan SJ, Carlbring P, Shafran R, Wade TD. Artificial intelligence as a virtual coach in a cognitive behavioural intervention for perfectionism in young people: A randomised feasibility trial. Internet Interventions. 2024;38:100795.

11.           Nicol G, Wang R, Graham S, Dodd S, Garbutt J. Chatbot-delivered cognitive behavioral therapy in adolescents with depression and anxiety during the COVID-19 pandemic: feasibility and acceptability study. JMIR Formative Research. 2022;6(11):e40242.

12.           Liu H, Peng H, Song X, Xu C, Zhang M. Using AI chatbots to provide self-help depression interventions for university students: a randomized trial of effectiveness. Internet interventions. 2022;27:100495.

13.           Qi Y. Pilot quasi-experimental research on the effectiveness of the Woebot AI chatbot for reducing mild depression symptoms among athletes. International Journal of Human–Computer Interaction. 2025;41(1):452-9.

14.           Romanovskyi O, Pidbutska N, Knysh AY. Elomia chatbot: the effectiveness of artificial intelligence in the fight for mental health. 2021.

15.           Shoshani A, Gurfinkel B, Kor A, Ben-Haim Y, Kanarek O, Segev R, et al. Efficacy of a Conversational AI Agent for Psychiatric Symptoms and Digital Therapeutic Alliance: A Randomized Clinical Trial. JAMA Network Open. 2026;9(4):e266713.

16.           Wu Y, Song H, Ye C, Lin R, Huang W, Wang Y, et al. A pilot randomized controlled trial of AI-delivered vs. human-delivered iCBT for depression in young adults. BMC psychiatry. 2026.

17.           Zhao Y, Qian W, Chen Y, Wu D, Luo Y, Gao C, et al. Effect of an AI agent trained on a large language model (LLM) as an intervention for depression and anxiety symptoms in young adults: A 28‐day randomized controlled trial. Applied Psychology: Health and Well‐Being. 2025;17(5):e70067.

18.           Huang L, Yu W, Ma W, Zhong W, Feng Z, Wang H, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems. 2025;43(2):1-55.

19.           National Mental Health Consumer Alliance. AI and digital mental health tools in Australia: risks, regulation and consumer leadership. 2026. https://nmhca.org.au/__static/jdj5jdewjgtjnnyzugvqodniljcztvbr/position-paper-ai-and-dmht-in-australia-final.pdf.

20.           Australian Government Department of Health DaA. Safe and responsible artificial intelligence in health care: legislation and regulation review final report. Canberra: Australian Government Department of Health, Disability and Ageing: Australian Government Department of Health DaA; 2025. https://www.health.gov.au/sites/default/files/2025-07/safe-and-responsible-artificial-intelligence-in-health-care-legislation-and-regulation-review-final-report.pdf.

21.           Goodman AG, Shinners, L, Mahoney, R, Victorian Aboriginal Community Controlled Health Organisation, The Aboriginal and Torres Strait Islander Community Health Service, The Centre of Excellence for Aboriginal Digital in Health and the Australian Indigenous HealthInfoNet,. Artificial intelligence for healthcare in Australian Indigenous communities: scoping project to explore relevance. CSIRO; 2025. https://www.csiro.au/en/news/All/News/2025/October/CSIRO-report-highlights-need-for-Indigenous-led-approach-to-AI-in-healthcare.

22.           Australian Institute of Health and Welfare. Digital mental health resources for First Nations people. AIHW AG; 2023. https://www.aihw.gov.au/getmedia/28f92ca3-ee7f-478a-828f-cfa2e00916e7/aihw-imh-20-digital-mental-health-resources-for-first-nations-people.pdf.

23.           Maidment K, Newby, J., Grant, H., Lattimore, J., Pollard, P., Scott, N. & Whitton, A. AI in mental health roundtable report. Sydney: Black Dog Institute; 2026. https://www.blackdoginstitute.org.au/wp-content/uploads/2026/02/BDI_AI-in-Mental-Health-Roundtable-9-Feb-2026.pdf 

24.           Rooney EJ, Makaza M, Wilson RL. Accepting the Legitimacy of Difference: Tools to Support the Decolonisation of Human Research Ethics in Western Health Research. Nursing Open. 2025;12(7):e70262.

25.           Esgin T, Macniven R, Crouch A, Martiniuk A. At the cultural interface: A systematic review of study characteristics and cultural integrity from twenty years of randomised controlled trials with Indigenous participants. Dialogues in Health. 2023;2:100097.

Authors:

  • Alicia Randell, Research Assistant, Knowledge Translation, Orygen, Centre for Youth Mental Health, University of Melbourne
  • Alan Bailey, Research Fellow, Knowledge Translation, Orygen, Centre for Youth Mental Health, University of Melbourne
  • Zoe Nikakis, Senior Project Officer, Knowledge Translation, Orygen
  • Dr Cali Bartholomeusz, Senior Academic Specialist, Knowledge Translation, Orygen, Centre for Youth Mental Health, University of Melbourne

Advisory Group members:

  • A/Prof Caroline Gao, Data Science & Analytical Methods (DSAM), Orygen
  • Dr David Baker, Strategy and Policy, Orygen
  • A/Prof Dom Dwyer, Dom’s Lab, Orygen
  • Dr Imogen Bell, Orygen Digital
  • Dr Isabelle Scott, Orygen Digital
  • Peta Humphreys, Youth Advisory Council (YAC), Orygen
  • Dr Shane Cross, Orygen Digital
  • Dr Shaunagh O’Sullivan, Orygen Digital
  • Steeley Shortland, Parkville Youth Mental Health and Wellbeing Service (PYMHWS) Specialist Program
  • William Cook, Digital Youth Advisory Group, Orygen

Authors gratefully acknowledge the contributions of the following people, whose expertise was instrumental in shaping this work:

  • Orygen National Youth Advisory Council
  • Orygen Youth Participation Team
  • Orygen’s specialist and primary youth mental health services family peer work program teams
  • Orygen First Nations Team