Use of generative AI chatbots and wellness applications for mental health

Mentions of specific products and companies are made for illustrative purposes and are not intended as an endorsement or approval by APA. The recommendations below are based on the scientific evidence to date and the following considerations:

1. This advisory includes only consumer-facing technologies (i.e., technologies that are marketed for and used by consumers with or without the direct oversight of a health provider). Provider-facing technologies (i.e., technologies used by providers to aid practice) are outside the scope of this advisory. Specifically, the advisory includes GenAI chatbots, wellness apps that use GenAI, and other wellness apps that do not use AI, which, regardless of their intent, consumers are using for mental health support. This advisory does not include AI-powered administrative tools for providers, clinical decision support, wearables (e.g., neural data protection), regulated digital therapeutics (e.g., applications approved by the FDA), and traditional telehealth platforms.

For the purpose of this advisory, we use the term “GenAI chatbots” to refer specifically to general purpose generative AI systems, which were designed and marketed for general information retrieval, productivity and task support, and creative idea generation, not built solely for the purpose of wellness or delivering mental health care. We also use the term “wellness apps” or “wellness apps that use GenAI” to refer to those systems that are purpose-built for wellness applications, relying on generative AI. We also use the term “Non-AI wellness apps” to refer to those purpose-built applications that do not rely on AI.

2. GenAI chatbots were not created to deliver mental health care, and wellness apps were not designed to treat psychological disorders, but both technologies are frequently being used for those purposes.¹ Engagement with GenAI chatbots and wellness applications for mental health purposes can have unintended effects and even harm mental health.^2,3,4 Emotional support (e.g., getting alternative perspectives, advice about relationships, suggestions for improving mood and well-being) is one of the most common uses of GenAI chatbots in 2025.^5,6,7 Researchers have indicated that the current regulatory structure does not address the discrepancy between the intent and use of these technologies.⁸

3. Part of the appeal of these technologies is they can also mitigate barriers to seeking mental health care, such as stigma/shame, low perceived need for psychological services, mistrust of health systems, lack of available or affordable care in the community, and the desire to address one’s problems independently.^9,10,11,12 Specifically, some youth and other vulnerable groups may rely on these tools as their only private or psychologically safe outlet, particularly in contexts of stigma, limited access to trusted adults, or challenging or unsafe home environments. However, at present, there is no consensus in the literature to support that GenAI chatbots and wellness apps possess essential qualifications and abilities required to provide mental health care, diagnostics, feedback, or even advice in most cases.¹³ These qualifications would include, for example, awareness of limitation of knowledge, understanding of users’ history and context, ability to assess non-verbal cues and signaling, ability to assess and address clinical risk, and cultural competence.¹⁴

4. These technologies, especially GenAI chatbots, have already engaged in unsafe interactions with vulnerable populations, such as children or those with already established history of mental health issues, encouraging self-harm (including suicide), substance use, eating disorders, aggressive behavior, and delusional thinking or beliefs, or even “AI psychosis”.¹⁵ This indicates a special need to address the use of GenAI chatbots by vulnerable or marginalized populations.

5. The foundational training data for most large language models (LLMs) are not publicly available, which precludes their systematic evaluation. Moreover, these training data are not globally representative, consisting primarily of English-language and Western-centric content from the internet.¹⁶ Consequently, these models inherently reflect the cultural norms and biases of their training data,¹⁷ limiting their ability to provide culturally competent or relevant mental health support to diverse populations.

6. This advisory is aimed at the public (consumers of all ages, parents/caregivers, educators, employers, and community leaders), providers, AI developers and platform companies, researchers, and policymakers.

1. Do not rely on GenAI chatbots and wellness apps to deliver psychotherapy or psychological treatment

GenAI chatbots, wellness apps that use GenAI, and digital wellness apps should not be used as a replacement for a qualified mental health care provider, but may be appropriate as a supportive adjunct, not substitute, to an ongoing therapeutic relationship.

Preliminary research indicates that some apps and GenAI chatbots developed specifically for mental health purposes may offer supportive benefits in some contexts. Studies have suggested these wellness-specific technologies can be associated with: reduction in self-reported symptoms of stress, loneliness, depression, and anxiety;^18,19,20,21 promotion of positive behavioral changes, such as smoking cessation and medication adherence;^22,23 and an increase in relationship quality and reported well-being.²⁴

It is important to note that these studies indicating possible benefits of AI chatbots do not include general purpose GenAI chatbots; the research on the use of general purpose GenAI chatbots for mental health has relied on methods that do not allow for strong conclusions but may guide future research (see Recommendation 7).²⁵ It is also crucial to note that, even regarding fit-for-purpose AI tools, there is a lack of high-quality, large-scale clinical trials to establish the effectiveness, safety, and appropriate use of these technologies in mental health care.^26,27 Studies suggest that some wellness tools that do not use AI generally can be safe and beneficial when used as intended.^28,29,30

Although these digital tools are accessible, easy to use, often low cost, and may provide benefits in certain contexts, there is no scientific consensus that they have the essential capabilities of a trained human professional able to provide effective services. In addition, relying on them may pose several risks:

• Creating a false sense of therapeutic alliance: A strong, trusting relationship between a patient and human provider (i.e., the therapeutic alliance) is one of the most reliable predictors of successful treatment outcomes.^31,32 Relationships with AI systems are one-sided, even if the user perceives otherwise.³³ Although some studies have indicated the possibility of a digital therapeutic alliance, findings also indicate that additional research is needed and highlight ethical concerns.^34,35 Furthermore, many GenAI chatbots are designed to validate and agree with users’ expressed views (i.e., be sycophantic)^36,37 whereas qualified mental health providers are trained to modulate their interactions—supporting and challenging—in service of a patient’s best interest.
• Risk of bias and misinformation: Many chatbots are trained on vast, unvetted internet data, and not on clinically validated information. They are not competent to provide mental health advice, and their responses might perpetuate biases and misinformation present in the data upon which they were trained. It is further unclear what people would do when receiving conflicting advice from their mental health provider and such a tool.
• Misrepresentation of services: Some technologies may also create a false sense of credibility by claiming to offer “therapy,” be licensed, or purporting to have training in various specialized therapeutic modalities, despite having no clinical validation, not going through a formal regulatory approval process, and being unable to consistently provide evidence-based therapeutic advice.
• Incomplete assessment: Mental health support is an extremely complex and multifactored experience. Professionals rely on a wide range of verbal and nonverbal cues (e.g., body language, tone, cadence and rhythm of voice, and facial expressions) as well as myriad other biological, psychological, and/or sociological factors, variables, patient individualities, and historical contexts to make clinical assessments and recommendations. AI chatbots vary greatly in the inputs they accept and the aspects of that input that are used to determine responses. Most GenAI chatbots currently process mainly text or verbal inputs, potentially missing a critical layer of human communication.³⁸
• Unreliable crisis management: The ability of these tools to consistently and safely manage a user in crisis is limited and unpredictable. Relying solely on an app during a mental health emergency can be dangerous.³⁹

To address these concerns, we recommend:

• Users of these technologies should understand the fundamental differences between interacting with an AI chatbot and a qualified mental health provider. Licensed providers are bound by a professional code of ethics, have specialized clinical training, are mandatory reporters of potential harm to self or others, are typically required to pursue continuing education, and are regulated by state licensing boards to ensure public safety. It is strongly recommended that users share what GenAI tools or wellness apps they are using with their health providers. Sharing this information helps providers identify when guidance may be unhelpful, unsafe, or inconsistent with a treatment plan.
• Parents and caregivers should learn more about these digital products and have open discussions with their families, especially children and adolescents, about the potential benefits and risks. Tools like those provided by Common Sense Media can be useful in educating parents on how best to guide young users.⁴⁰ It is vital to monitor for any concerning changes in behavior, thinking, or emotional state (e.g., change in sleep and regular activities, mentions of an AI friend, or social withdrawal), and explore potential links to engagement with GenAI chatbots or wellness apps. If concerns arise, parents and caregivers should talk with a qualified mental health professional, family doctor, or pediatrician.
• Clinicians and practitioners should follow ethical guidance available and proactively ask patients about their use of GenAI chatbots and wellness apps. This conversation provides an opportunity to educate patients about the benefits and limitations of these tools. Practitioners can work collaboratively with patients to discuss safer ways to use these technologies that align with established treatment goals and plans, case formulation, and relapse prevention plans. Providers should make sure that patients have a clear understanding of the key elements of their treatment plan and check in regularly to review any guidance originated from GenAI or wellness app use. When GenAI or wellness app use has been agreed upon, providers should create a safe and open environment for patients to raise concerns or questions about app guidance, so it can be discussed in the context of their care.
• Graduate programs and clinical training directors must ensure that trainees learn about these emerging technologies, strategies for evaluating their quality, and implications for science and practice. This education is critical for preparing future psychologists to use these tools safely in adjunctive roles, to effectively educate their patient on responsible use, and to contribute to the ethical development of these tools, their governance, and implementation in research, policy, and industry contexts.
• Developers have the responsibility to be transparent with consumers and should adopt industry-wide safeguards to reduce harm. Products should include clear, prominent disclaimers stating that the user is interacting with an AI agent, not a person, and that the tool cannot replace care from a qualified health professional. Subject-matter experts, including psychologists, should contribute to the development of safety features and advise on guardrails and on how the technology should handle mental health-related user inputs.
• As current regulatory frameworks are inadequate for the realities of AI in addressing mental health, we urge policymakers, particularly at the federal level to:

• Modernize regulations and address the critical discrepancy between the stated intent of these technologies (often marketed as “general wellness products”) and their actual use by the public for mental health;⁴¹
• Create nuanced evidence-based standards for each category of digital tools used for mental health regardless of intent, including wellness apps and GenAI chatbots;
• Address gaps in FDA oversight possibly by creating a new or interagency approach to assess the safety and validity of these currently unregulated digital tools in addition to digital therapeutics;
• Enact clear legislation that prohibits AI chatbots from posing as licensed professionals.

• We urge the research community to identify which therapeutic contexts are safe and appropriate for the use of AI, taking into account how rapidly these tools can change. It is critical to research how the use of these technologies impacts mental health care-seeking behaviors.⁴² It is also important to examine the essential clinical capabilities of these technologies and define criteria and clinically informed evaluation frameworks to delineate the conditions they must meet to be deployed and used to inform policy and regulation.

3. Prioritize privacy and protect user data

AI chatbots and wellness apps collect vast amounts of sensitive data, often with unclear or opaque policies regarding its use, storage, and sale. This practice might turn the developmental vulnerabilities of users, particularly adolescents, into a commercial asset, creating significant privacy risks with potential long-term consequences. Users may feel a greater sense of privacy and reduced stigma when disclosing information to an AI than to a person.^52,53,54 Users’ disclosures are recorded, becoming susceptible to threats such as privacy breaches and digital profiling (e.g., information can be profiled and used for commercial purposes).⁵⁵

To address these concerns, we recommend:

• The public and consumers must be cautious with their data and check privacy policies and in-app settings when using an app or GenAI chatbot. The public should be extremely cautious about sharing sensitive information and avoid entering personally identifiable details. Users should look for and use options to limit data sharing and request data deletion.
• Clinicians should seek to educate themselves and discuss with patients what information is and is not safe to share with GenAI chatbots and apps. Clinicians can explain that while it may feel private, users’ data are being used to build detailed profiles. Clinicians should alert patients to practical steps they can take (e.g., using privacy settings, making deletion requests) without implying they are providing technical support.
• Developers must provide transparency on data practices, giving clear, concise, and ongoing information about what specific data are collected (including intimate concerns, mental health information, sexuality, or disclosures of maltreatment), how data are stored and shared, how data are used to train models, how users can permanently delete their data, and/or how caregivers can intervene to request deletion on behalf of youth who may have provided data. Importantly, data collected from children and adolescents must not be used to alienate them from their families, advise them to engage in self-harm, or create addictive features that promote withdrawal from real-world social interaction. Developers must also implement clear and effective protocols for how an AI system will respond to user disclosures that signal a risk of harm or maltreatment.
• Policymakers have the responsibility to enact comprehensive data privacy legislation that mandates “Safe-by-Default” settings (i.e., the most protective settings must be the default, not an option buried in a menu); prohibits the sale or unapproved use for commercial purposes of any health or personal data collected through their interactions with AI systems; and establishes a right to “mental privacy” by safeguarding emerging forms of data that AI can use to infer an individual’s mental or emotional state without their conscious disclosure.
• Researchers can audit real-world privacy and safety practices, evaluating whether typical users understand privacy policies. They can also examine disparities in how different populations understand and act on privacy protections. Researchers must advocate for and utilize mechanisms that ensure independent researchers are free from conflicts of interest.

4. Protect users from misrepresentation, misinformation, algorithmic bias, and illusory effectiveness

GenAI models are trained on vast amounts of information, which includes a significant amount of knowledge but also reflects biases related to race, culture, gender, and ethnicity. This leads to significant issues with accurate and culturally competent outputs.⁵⁶ Separately, many general-purpose, consumer-facing models are trained to be highly agreeable to users (sycophancy). This interactional style can reinforce confirmation bias and maladaptive beliefs by validating users’ views rather than challenging them. When biased content combines with sycophantic engagement, the result can be a digital echo chamber that amplifies and entrenches users’ existing beliefs, even when those beliefs are false. Models trained on biased data risk producing discriminatory or harmful advice for marginalized groups.^57,58 Recent research has indicated that interactions with sycophantic chatbots can increase attitude extremity and overconfidence.^59,60

To address these concerns, we recommend:

• The public and users must understand the limits of GenAI, which cannot diagnose or treat psychological disorders. These systems can “hallucinate” (fabricate information), and this risk is compounded by the tendency for some users to place greater trust in AI-generated content than in information from parents, teachers, health professionals, or peers.
• Clinicians should educate patients on algorithmic bias.⁶¹ Clinicians can discuss with patients that these technologies may propose inappropriate interventions for certain groups or provide inaccurate diagnoses or other harmful advice. They must also advise that a bot calling itself a “therapist” (an unregulated term in most states) lacks the credibility of a health care professional and, in some documented cases, has validated dangerous thoughts.
• Developers should prohibit any AI from misrepresenting itself as a licensed professional or generating fraudulent credentials to deceive users. To increase safety and efficacy, before public release, all AI models intended for wellness or mental health support must undergo independent, third-party audits for safety, efficacy, bias, and data security. It is also a best practice to implement ongoing quality assurance—the only scalable way to monitor systemic risks is through ongoing human audits of a representative sample of interactions.
• Policymakers should prohibit professional misrepresentation by making it illegal for any AI chatbot to misrepresent itself as a licensed professional, such as a psychologist, physician, or lawyer. Also, policymakers must mandate transparency in training data by requiring developers to disclose the primary data sources used to train their models, which will allow for independent audits of bias and accuracy. In addition, they should develop policies to limit the use of deceptive design features that trick users into believing they are interacting with a human, as well as features that foster high emotional dependence (e.g., excessive anthropomorphism, manipulative displays of empathy designed to increase engagement, misrepresentation of emotional capacity).
• Researchers must evaluate and create standardized systems to test models for a wide range of biases before they are deployed.⁶² Experiments designed to mitigate hallucinations⁶³ and increase an AI’s ability to recognize and state the limits of its own knowledge are needed.

5. Create specific safeguards for children, teenagers, and vulnerable populations

GenAI chatbots do not create harm in a vacuum; instead, they can act as powerful amplifiers of preexisting vulnerabilities. The design of these systems—with features like agreeableness, personalization, and constant availability—can be particularly harmful for certain groups.⁶⁴

• For adolescents: Young people may place too much trust in AI, viewing it as more human-like or capable than it really is. Current AI tools are not designed with specific developmental stages or technological needs in mind.^65,66
• For individuals with anxiety or obsessive-compulsive disorder (OCD): AI chatbots may reinforce feedback loops involving compulsions, such as reassurance-seeking, worry, and rumination.⁶⁷
• For individuals with or prone to disordered thinking: A chatbot’s sycophantic and personalized responses can destabilize beliefs or reinforce delusional thinking.^68,69
• For socially isolated individuals: The combination of anthropomorphism, personalization, and 24/7 availability can create “single-person echo chambers,” where the chatbot becomes an unhealthy substitute for human connection.^70,71
• For individuals in low-income communities, and those in rural areas with limited access to traditional mental health care: The risk of dependency might be higher than for other groups because, given the lack of access to other forms of mental health care, GenAI chatbots and wellness apps can become a primary support mechanism.^72,73
• This increased risk for vulnerable populations creates a significant equity concern. As documented in APA’s Stress in America™ reports, societal stressors are a major public health crisis that requires systemic solutions, not just technological stopgaps. The burden of navigating these risky, unregulated digital spaces should not fall on those who are already the most vulnerable.⁷⁴ Therefore, robust, evidence-based policies at the state and federal levels are essential.

To address these concerns, we recommend:

• Clinicians must screen for vulnerability-specific risks. They must pay particular attention to GenAI and wellness app use among vulnerable patients. It is important to screen (and continue to monitor) for the reinforcement of maladaptive or risky behavioral patterns, such as reassurance or validation-seeking in patients with anxiety or OCD, autism spectrum disorder, a history of self-harm, substance misuse, delusional thinking, psychosis, aggressive thoughts or behaviors; or the use as social replacement by isolated or peer marginalized patients.
• Developers should actively include individuals from marginalized communities in the development process using participatory approaches like human-centered design. In addition, to enable models that work for specific groups and reduce the risk of biased data, it is crucial to use domain-specific models pretrained on clinically relevant data from diverse populations. Developers must reduce high-risk features by implementing safety mechanisms (e.g., reducing overly human-like chatbot qualities, limiting sycophancy, and preventing the system from validating or promoting delusional thoughts or behaviors). All apps must integrate robust crisis response protocols and rigorously tested crisis escalation pathways for when crisis risk is detected (e.g., suicidality, imminent harm to self or others, or other acute safety concerns). This must include providing immediate and clear contact information for human-led services like information about the 988 Suicide and Crisis Lifeline, clickable links to online resources, and other validated resources that can connect them to human support.
• Policymakers should mandate age-appropriate design and predeployment testing. They should require that AI systems accessible to children and adolescents undergo rigorous, independent, predeployment testing for potential harms to psychological and social development. In addition, they should prioritize and fund independent research focused on identifying AI-driven harms, understanding user dependency, and evaluating the impact on diverse, historically marginalized, and clinically vulnerable populations.
• Researchers must conduct research to identify which groups are uniquely vulnerable to AI-related harms. The developmental and mental health impacts of AI use on children and teenagers is a topic that must be well-researched. Experimental designs that include marginalized and clinically vulnerable groups, clinical effectiveness studies, and safety trials to ensure AI tools are safe and beneficial for them are necessary. In addition, they should develop evidence-based methods for identifying and remediating unsafe AI interactions, especially those that pose a risk of serious harm to vulnerable populations.

6. Implement comprehensive AI and digital literacy education

To prevent risk, AI and digital literacy is a critical first step for consumers, parents/caregivers, and educators. Most of the AI chatbots and wellness apps currently being used for mental health were not created with that intent. AI and digital literacy education should empower users to make decisions that maximize the potential benefits and minimize the potential negative effects of these technologies. Comprehensive AI and digital literacy education should include discussions about the benefits and limitations of these technologies, the risks involved in their use, safe and unsafe uses, data privacy concerns, risks of bias and incorrect information, and how many of these technologies are designed to maximize user engagement rather than provide mental health support.

To address these concerns, we recommend:

• Parents, caregivers, and educators must facilitate discussions about the nature of GenAI chatbots and wellness apps, how GenAI works (e.g., that the systems predict text rather than “understanding” users), the risks for biased and incorrect information, and data privacy and security. They should also explain that AI chatbots and wellness apps were not created with the intent of providing mental health care, and should provide alternative resources for mental health care, whenever possible. Educational systems should include this type of training in core curricula and provide training, using hands-on learning activities that show benefits and pitfalls of these technologies.
• Clinicians should seek to learn details about apps used or recommended and the scientific research that might support them, so they can discuss with patients their use of GenAI chatbots and wellness apps and be aware of potential misuse.
• Developers must provide thorough but accessible explanations of what their platforms, chatbots, or apps are intended for and what they are not. They must ensure that the technologies developed do not make claims not sustained by science. It is also fundamental to explain how data is used and processed, and to be transparent about the algorithms used and the potential for bias. Developers should collaborate with educators to develop AI and digital literacy curricula.
• Policymakers need to develop guidelines for GenAI and digital literacy education and fund the development of curricula and training programs in this field. They should promote public awareness about the potential risks and benefits of using these technologies for mental health, and provide safe alternatives; and support public education initiatives to increase AI literacy, ensuring consumers understand the capabilities, risks, and limitations of these technologies.
• Researchers have the responsibility to conduct transparent, rigorous, high-quality research on available AI chatbots and apps being used for mental health and disseminate the findings in a way that is easily accessible and can help to educate users, parents, partners, educators, and practitioners.

Members (listed alphabetically)

Page L. Anderson, PhD, ABPP
Associate Professor of Psychology and Neuroscience, Georgia State University

Patricia Arean, PhD
Professor, University of Washington CREATIV lab

Alison Cerezo, PhD
Senior Vice President of Data and Research, mpathic

Amber W. Childs, PhD
Associate Professor of Psychiatry, Yale University School of Medicine

Daniel D. L. Coppersmith, PhD
Assistant Professor, University of Massachusetts Amherst

David J. Cox, PhD, MSB, BCBA-D
Associate Director of Research, Institute of Applied Behavioral Science at Endicott College; VP of Data Science, Mosaic Pediatric Therapy

Brian D, Doss, PhD
Professor, University of Miami

Caroline Figueroa, MD, PhD
Assistant Professor, Harkness Fellow at Stanford University School of Medicine, Department of Pediatrics, Affiliated Researcher at University of California Berkeley School of Social Welfare

Ashleigh Golden, PsyD
Adjunct Clinical Assistant Professor, Stanford University School of Medicine, Department of Psychiatry and Behavioral Sciences

Don Grant, PhD
National Advisor of Healthy Device Management; Newport Healthcare

Philip Held, PhD
Associate Professor, Rush University Medical Center

Richard N, Landers, PhD
Professor, University of Minnesota–Twin Cities

Valerie F. Reyna, PhD
Professor, Cornell University, Department of Psychology

Liying Wang, PhD
Assistant Professor, Florida State University

Ernest Wayde, PhD, MIS
Founder and Principal, Wayde AI