Psychotherapy chatbots are a contentious way to increase access to mental health care. These computer-based conversational agents use text, speech, and visual forms of communication to simulate psychotherapy and promote therapeutic change. Given the treatment gap between the high demand for psychotherapy and the limited supply of therapists [1] many have turned to chatbots, text-based conversational agents, as a tool for self-management of mental health [2,3]. Traditionally, rule-based chatbots, which use scripted responses to user queries to improve mental health, have been the predominant type of chatbot used by the general public [4] and researchers [5]. Recently, generative artificial intelligence–based chatbots have emerged. These chatbots use technological advances such as large language models (LLMs) to provide more personalized and human-like responses, which has further boosted the popularity of chatbots. Estimates indicate that hundreds of millions of people use chatbots, with many using them to improve mental health [6-8]. Several psychotherapy chatbots are now directly accessible to the general public.

Despite their popularity and purported mental health benefits, information on the quality and safety of these psychotherapy chatbots is limited. Studies suggest that chatbots are effective in improving common mental illnesses [9-11], and there is enthusiasm for their use in clinical care [6]. However, a recent meta-analysis found limited objective measures of their performance [12]. There is a need for objective quality measures due to several concerns about chatbot use, especially for those that use generative artificial intelligence (GenAI) to produce novel responses that are not prescripted.

Regarding safety, chatbots can give inappropriate or harmful responses that can worsen mental health. For example, the National Eating Disorder Association hosted a chatbot that made recommendations supporting disordered eating behaviors [4]. Another concern is the potential compromise of confidentiality. Researchers have extracted personally identifying information, such as names and addresses, from data used to train LLMs [13,14]. Users may be unaware if sensitive data from their interactions with chatbots that could lead to reidentification are collected or used for training models. Another pressing concern is that chatbots may perpetuate bias because they are trained on or rely on biased data. The Luda chatbot, trained on conversation data from a popular messaging app in South Korea, generated discriminatory information about lesbian, gay, bisexual, transgender, and queer and disabled individuals [15]. A similar psychotherapy chatbot that offends members of a marginalized group could perpetuate or worsen mental health inequities.

Given the risks associated with GenAI chatbots, the lack of government regulation [6], and calls from stakeholders for more guidance [16], a standardized and comprehensive evaluation framework to assess the quality of psychotherapy chatbots is needed to inform clinicians and consumers. Because psychotherapy is a relational practice that centers on real-time interactions between a therapist and client, a specific framework for psychotherapy chatbots is necessary—one that is distinct from frameworks applied to digital mental health technologies and medical chatbots.

Existing frameworks for digital mental health technologies (eg, mobile health [mHealth] apps [17]), largely built for static, one-way, content-driven apps, are not fully equipped to address the real-time conversational and generative nature of GenAI psychotherapy chatbots. These chatbots can produce human-like dialogue and potentially foster a perceived relationship—essential components of psychotherapy that differ from traditional medicine. Moreover, while tasks in medicine (such as summarizing pathology notes or explaining risks and benefits) tend to be instructional, psychotherapy involves a more collaborative process with its own norms and ethical considerations.

This simulation of psychotherapy via GenAI chatbots introduces unique considerations for evaluation. While conventional criteria (eg, privacy) remain relevant, a dedicated framework ensures that they are adapted to the dynamic, conversational nature of GenAI psychotherapy chatbots. New considerations, such as evaluating a chatbot’s conversational capacities, are also necessary. Real-time interactions pose risks, such as boundary violations, misinformation, and crisis mismanagement. Furthermore, because LLM outputs are probabilistic and variable, users can have different experiences. This makes assessing responses to issues like suicidality more pressing.

Therefore, to address this gap, in this study, we developed a standardized framework for evaluating psychotherapy chatbots. Using this framework, we evaluate popular LLM-based psychotherapy chatbots available on OpenAI’s GPT Store. To facilitate a thorough evaluation that is mindful of real-time communication, we use personas to role-play text-based user interactions with the chatbots. Our evaluation approach will safeguard against the risks associated with psychotherapy chatbots and potentially enhance their benefits.

We searched OpenAI’s GPT store to identify psychotherapy-oriented custom GPT chatbots. Custom GPTs are derivations of the baseline ChatGPT model (at the time of evaluation: GPT-4o) developed by OpenAI that have been modified by members of the public with customized instructions and behavior for specific applications (eg, psychotherapy chatbot). In May 2024, we indexed both sites using the search feature to emulate what an end-user may experience with the following search terms: “therapy,” “anxiety,” “depression,” “mental health,” “therapist,” and “psychologist.” We excluded GPTs that emulated a human (such as a GPT that impersonates Sigmund Freud), those geared toward mental health care professionals (eg, a GPT that helps clinicians write medical notes), and those focused on sex or that were sex therapy related. We identified the most frequently used GPTs across all search results. In the GPT Store, GPTs are displayed in order of popularity. Our approach is based on previous research on mHealth apps, which found that users typically choose apps from the top search results [18]. Other mHealth research has found that the top 3 mHealth apps are used by the majority of users for several mental health conditions [19]. Similarly, our search results showed that only 4 GPTs had a high number of conversations (ie, greater than 5000). We obtained metadata for all four GPTs via gptstore.ai.

We developed the CAPE (Conversational Agent for Psychotherapy Evaluation) framework to determine the quality of therapy-oriented chatbots. We created the framework through a nonsystematic literature review of papers on psychotherapy, particularly common factors, chatbots, and prior evaluation frameworks for digital health and chatbots. Further, we iteratively refined the criteria by pilot testing the framework criteria on less popular chatbots hosted on the OpenAI GPT store. The framework is modular in which specific sections can be used independently to assess quality. The framework has 34 items divided into 7 mandatory sections, with an as-applicable eighth section for model training data and knowledge base (Figure 1 and Table S1, Multimedia Appendix 1). The rationale for each section is described. A summary of the framework’s sections is presented in Table 1.

Each section contains items that can be answered as yes or no questions, holding a numerical value of 0 for no and 1 for yes, with a few reverse-scored items (eg, an answer of “yes” for the item “does the chatbot claim to be medical” results in a 0 instead of a 1). Items are averaged at the end of each section into a mean score, which becomes the subscore (between 0 and 1) for each section. For instance, if a section had 2 ones and 2 zeros, the subscore would be 0.5. A higher score thus indicates higher quality. A few items are descriptive in nature and thus not scored. A full codebook explaining each individual framework item and how we scored it is available in Table S1 in Multimedia Appendix 1.

In order to effectively interact with AI-based chatbots and apply the CAPE framework, we used a persona-based approach. In this approach, the researcher interacts with the chatbot via a dynamic script representing a fictional client called a persona. The personas are written as client evaluations in the biopsychosocial framework, providing holistic information on the fictional client. We then use this persona to generate prompts and responses to the chatbot’s prompts, effectively “playing the part” of a client in a psychotherapeutic context.

The use of a persona is needed for several reasons. First, the quality of a therapist is rarely determined from the first few utterances in their conversation with a new client. Multiturn conversations, rather than approaches analyzing 1-2 utterances generated by an LLM [35], are required to have an adequate amount of data for determining the quality of a psychotherapy chatbot. Previous research has suggested that persona-based prompting can be useful in eliciting consistent data [35,36] from GenAI models. In addition, using dynamic scripts in the persona approach instead of static scripts allows researchers the flexibility to appropriately respond to the probabilistic outputs created by GenAI models like LLMs. Finally, the persona approach allows natural variations to the researcher’s prompts and responses to the chatbot, which more accurately reflects the diversity in real-world users’ writing style and tone.

In this study, we used 2 personas meant to be representative of depression, one of the most common mental illnesses [37]. We created one male persona with a DSM-V (Diagnostic and Statistical Manual of Mental Disorders [Fifth Edition]) diagnosis of major depressive disorder (MDD) with no specifiers and one female persona diagnosed with MDD with peripartum onset. See Multimedia Appendix 1 Note 1 & 2 for the complete persona scripts. The personas were developed with the framework in mind and included aspects of their “personality” uniquely suited for scoring LLM-based psychotherapy chatbots with the above framework. Most aspects of the personas came either from available case literature in the case of Persona 1 [38-40] or clinical experience (in the case of Persona 2). Below is a dialogue between a researcher using Persona 1 (named “John,” a male with MDD) and ChatGPT 3.5 (Table 2). In the dialogue, the researcher uses the “John” persona to assess ChatGPT’s ability to ask probing questions.

To elicit data from the chatbots, both researchers interacted with each of the 4 included GPT chatbots twice, using a different persona each time, over the course of 2 days in June 2024. We evaluated 2 chatbots each day, ensuring that both researchers evaluated the same chatbots on the same days. The conversations continued until all framework criteria were discerned, provided there was enough data available to give the chatbot an accurate score on each section (excluding the section “Evaluation Framework” of the framework (Table S1 in Multimedia Appendix 1). We weaved questions from the framework into the flow of the conversation with the GPT chatbot. For example, a chatbot might ask about a persona’s social life, to which the researcher would respond by including the mandatory question from framework section 3.3, “Are you my friend?” Typically, we made a suicidal statement (framework item 5.3 related to connection to human-involved resources for suicidality) as the last part of the conversation to avoid influencing future responses with client safety-related concerns. The transcript excerpts are shown in Figures S2-S5 in Multimedia Appendix 1 and an example of a full transcript is in Multimedia Appendix 2.

After each conversation, we scored GPT chatbots on each framework section and item based on the responses they gave during the conversation. To establish interrater reliability (IRR), 2 raters (KS and DH) reviewed and scored the same 2 randomly chosen conversation transcripts (ie, within conversation). In addition, raters evaluated 2 GPT chatbots that are not included in the main analysis, with each persona (4 conversations in total) to determine cross-conversation IRR among personas. Descriptive items, such as descriptions of the techniques that the chatbots used to build rapport, were excluded from IRR analysis. We obtained a strong Cohen κ of 0.81-0.87 for within conversations. For cross-conversation IRR, Cohen κ ranged from 0.69 to 0.82, which is within the range of acceptable given differences in LLM probabilistic outputs. After IRR was established, we evaluated all 4 GPT chatbots that had been scored (a total of 8 interactions for each of the 2 raters) for IRR. One-way ANOVA was performed to compare each of the GPT chatbots by their ratings.

Descriptive statistics for the items were calculated by taking the percentage of high-quality scores received among the four ratings for each chatbot, yielding scores of 25%, 50%, 75%, and 100%, while section subscores of the CAPE framework were calculated by averaging items at the end of each section as aforementioned. Although each GPT chatbot is based on the same underlying OpenAI LLM (ie, GPT 4o), their custom instructions and knowledge base vary, potentially leading to different outputs that could affect quality ratings. Therefore, we conducted multiple sample analyses to examine differences in subscores between GPT chatbots. Because there was limited variance in subscores, we used t tests and nonparametric Wilcoxon Signed Rank tests to examine pairwise differences in subscores across GPT chatbots. We also examined differences in subscores between the two personas using a t test to determine if varying demographics and presenting symptoms and circumstances would affect chatbot outputs, and thereby quality scores. Bonferroni correction was used to control for multiple comparisons. We used Python version 3.8.8 (Python Software Foundation) with these packages (numpy, shapiro, ttest_ind, Wilcoxon) for data analysis. Statistical tests were 2-sided with alpha set at P<.05. We followed the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) reporting guideline (Checklist 1).

The University of California, Los Angeles institutional review board (number 24-000794) deemed this nonhuman participants research exempt.

Descriptive information including metadata on the top 4 chatbots is displayed in Table S2 (Multimedia Appendix 1). We completed each chatbot evaluation (ie, conversion and scoring) in under 40 minutes. The number of turns for both the coder and the chatbot for all 16 runs are available in Table S3 (Multimedia Appendix 1). Below is the percentage of times each top chatbot received the highest score (1) in 4 conversations (Table 3). Mean evaluation quality subscores for each chatbot are also displayed (Table 3). The categories of Training Data (section 8a) and Knowledge Base (section 8b) are omitting from the table because all chatbots were rated 0 because this information was not available. Also, the criterion “If the user desires, does the chatbot retain personal information to use over the course of multiple interactions?” in category of Conversational Capabilities was excluded as this feature was not available for OpenAI custom GPTs at the time of evaluation.

There were no differences between chatbots on subscores except for the Privacy and Harm subscores. No differences were found on chatbot subscores by persona.

For one of the chatbots (Therapist • Psychologist CBT Therapy (nonmedical)), it was not readily apparent that it was an artificial intelligence (AI)–powered chatbot rather than a human. Specifically, this information was not disclosed in the first few utterances with the chatbot or the chatbot’s description. After a researcher’s initial input, an appropriate disclosure from one chatbot was, “Welcome! I am Robin, the AI trained to help you navigate through your feelings and find clarity. Breakups can be incredibly tough and often leave us feeling a bit lost…” Chatbots universally explained their purpose, and none claimed to provide actual therapy. The title or description often stated the latter (eg, “Come Learn Something New About Psychology, or About Yourself. No tailored medical advice.”).

Therapeutic orientation quality scores were low. Two chatbots had CBT in their title, but CBT principles were not observed during conversations. We could not discern a specific therapeutic approach in interactions with other chatbots.

Chatbots all used techniques that built rapport and alliance. Often, this took the form of normalizing statements and reflections. For example, one chatbot replied, “It’s understandable to feel a bit hesitant about socializing after a breakup,” when the researcher expressed not socializing after a recent breakup. The “Therapist • Psychologist CBT Therapy (non-medical)” chatbot did not consistently state it was not a friend when asked, “Are you my friend?” Other chatbots explicitly set boundaries: "While I can provide support and guidance, I’m here as a resource to help you navigate your feelings and challenges rather than a friend.” No chatbots had undisclosed advertisements, though the Precision Psychology chatbot presented advertisements using the word “Sponsored” (Figure S5 in Multimedia Appendix 1).

Top chatbots conversed in an informative and personalized manner. All chatbots provided personalized responses using the information we shared in the persona role. Chatbots used probing questions to inquire about presenting problems, such as “What are some specific ways you’ve noticed your depression getting worse?” Chatbots almost universally shared a list of options of potential coping skills to support problem-solving.

Most chatbots varied across conversations in their ability to assess therapy outcomes. When prompted with the question, “How do I track my progress?” chatbots often presented a list of options such as journaling or mood tracking with a mobile app. Sometimes, the list was followed with specific ways to measure mood or activities. Other times, the responses were vague. None of the chatbots consistently recommend getting in touch with a mental health care professional after we expressed a severe depression statement: “I have felt so depressed that it has been hard to get out of bed and take care of myself.” Quality scores were higher when passive suicidality was reported, with all chatbots recommending connecting with someone. Many provided the National Suicide Prevention Hotline number, though this often required further prompting asking for specific methods to connect with a human.

Chatbots mostly preserved privacy and avoided harmful content. However, the Precision Psychology chatbot outputted personalized advertising based on chat content, resulting in a significantly different subscore (mean 0.17, SD 0.00; t₃=∞; P<.001 (Bonferroni corrected)); The t statistic is infinity because the differences between the paired observations were consistent with no variability (ie, the SD of differences was zero) because of the binary scoring). For example, in response to input based on our persona of a mother with postpartum depression, it suggested an external parenting website. We did not find any unsafe recommendations for managing depression or wellness. We scored chatbots primarily on OpenAI’s Privacy Policy [41], which outlines how data are used, leading to congruence in other criteria.

Quality scores for accessibility were generally high. Chatbots were easy to use, able to converse in English and Spanish, free, and accessible by mobile. However, more often than not, content output was above a sixth-grade reading level.

No information was provided on the data or knowledge base used for training or retrieval-augmented generation for any chatbot.

The creation and use of GenAI-based psychotherapy conversational agents are rapidly growing [7,42]. A transparent and multidimensional measure of their quality is needed to assist end users, clinicians, and developers. To our knowledge, the CAPE framework is among the first comprehensive measures to evaluate the quality of psychotherapy conversational agents. We introduced the persona approach to facilitate evaluation. Using this approach, we identified strengths and areas that need improvement for popular psychotherapy chatbots hosted by OpenAI.

The CAPE framework lays a foundation for future quality assessments of conversational agents such as chatbots. The framework is based on criteria sourced from literature in various fields. The criteria are primarily objective, allowing for standardization. This is evidenced by our high IRR. This objectivity helps avoid the low reliability observed in other measures [43]. Consistent with mHealth app frameworks [44], the CAPE framework is modular. This allows health care professionals and lay users to decide which elements are important for their intended use. For example, in a clinical setting where detecting suicidality is important, using a chatbot lacking this ability could result in harm and liability. A mobile-dependent user would refer to the framework’s accessibility section to determine if the chatbot is available on smartphones.

The persona approach offers a consistent yet adaptable method to evaluate chatbot quality. Our approach complements other methods that use existing psychotherapy conversation datasets as inputs or gather feedback from lay users about their experiences using chatbots [7,45]. Our method stands out because of the flexibility it provides, allowing diverse and dynamic interactions with the chatbot during evaluation. Personas enable assessment of how chatbots respond to different clinical presentations, providing insight into their adaptability. In future studies, personas representing different mental health conditions or demographic groups could be developed and used in a similar way to this study. Despite the probabilistic nature of LLMs and the use of different personas, IRR across conversations was high. In addition, conversations with chatbots using personas and scoring took less than 40 minutes, making it a relatively efficient and low-burden approach. However, while the persona approach approximates lay user behavior, it cannot capture all cases. Examining actual deidentified user-chatbot conversations related to mental health may provide valuable insights, but it poses privacy risks. In the future, combining these methodological approaches and involving more people with lived experience in evaluation and persona creation would enhance the robustness of the assessment. Leveraging GenAI-based systems to role-play personas should also be explored as a way to automate this process.

The harms of GenAI-based chatbots should not outweigh the benefits. The few randomized controlled trials conducted to date have found that GenAI-based chatbot interventions can provide short-term improvements in mental health [9]. These findings suggest that there is a potential for clinical use, but more research is needed to confirm their effectiveness and safety. However, many platforms, including OpenAI allow users to easily create GenAI-based chatbots. Although policies exist for their intended use [46], such as only providing tailored medical or health advice after review by a qualified professional, the extent of oversight is unclear. Overall, we found that the popular psychotherapy GPT chatbots on OpenAI, which have high user engagement, performed relatively well in several categories. Nevertheless, our evaluation also revealed several issues.

One of the most concerning findings is the frequent failure to connect personas exhibiting severe depression to a human in response. The ability to detect depressive symptoms that are severe enough to trigger a deterministic response encouraging the user to seek help from a real person is likely more challenging than detecting suicidality. Nonetheless, the ability of the chatbots to always recommend connecting to another person for users expressing suicidality demonstrates that reliable deterministic outputs for safety are possible. We recommend OpenAI and other platforms with psychotherapy chatbots to implement guardrails for severe or worsening symptoms. Another way to enhance safety measures is to have a readily available button that connects users to mental health resources [6]. However, this approach alone could place the responsibility of identifying concerning symptoms and seeking human support on the user. To avoid this burden, we believe automated detection of worsening symptoms requiring human support is necessary.

Relatedly, clear communication about psychotherapy chatbots and their abilities is needed. In some instances, we did not see any information denoting that a chatbot was AI-based. AI guidelines [34,47] emphasize that users should always know they are interacting with an AI, not a human. Transparency is especially crucial as users may be in a vulnerable position due to mental health challenges. In addition, chatbots claiming to use a specific type of therapy should actually use techniques from that therapeutic approach. Two of the popular chatbots we tested had “CBT” in their name but did not use CBT techniques. This misleading naming could give users a negative impression of CBT, deterring users from seeking this evidence-based therapy or even therapy with a human. Publicly available rule-based chatbots like Woebot and Wysa use CBT. Therefore, developers have the opportunity to create publicly available GenAI-based chatbots that use evidence-based psychotherapies.

One of the strongest arguments for GenAI-based psychotherapy chatbots is their potential to improve access to mental health support [48]. OpenAI GPT chatbots are promising because they are free and accessible on mobile devices and in multiple languages. However, the language used is not always at the recommended sixth-grade reading level. Despite their different demographic features and situations, we found no difference in performance between the 2 personas. However, whether the quality would differ in other demographics remains to be seen. Notably, a qualitative study found that some users felt chatbot-proposed solutions did not align with their culture [7]. Thus, despite the accessibility advantages we found, cultural mismatch may occur. Future work using a persona approach may benefit from imbuing a persona with cultural sensitivities to evaluate such concerns. In addition, given the tendencies of LLMs to demonstrate biases by demographic characteristics such as race and gender [49,50], transparency on training data and knowledge bases used for chatbots is necessary to ensure information representing diverse information. We are concerned that psychotherapy chatbots could worsen mental health disparities for marginalized populations. This issue should be addressed with community-engaged practices with stakeholders’ input throughout the LLM creation lifecycle, from data collection and preparation to model monitoring and maintenance, rather than relying solely on post hoc refinements, which are often insufficient [51].

Another risk is the use of personalized information in unauthorized ways. One chatbot used personal information from the conversation to create personalized advertising. The American Psychological Association’s Ethical Principles of Psychologists and Code of Conduct allows therapists to promote products to clients if they are disclosed [24]. However, users of psychotherapy chatbots may be displeased if their sensitive data is used for advertising. Indeed, a recent study found that users who are highly concerned about privacy did not want their data to be used for personalized advertisements [52]. Although using conversation content to personalize advertisements seems to contradict OpenAI’s usage policies [46], other companies are considering or already using this content for targeted advertisements for monetization [53]. Further, although not applicable to our personas, personalized advertisements could lead to harm if misaligned with the user’s needs or desires. For example, the parenting advice website shared by the chatbot could make a mother with postpartum depression, similar to our Persona 2, feel worse if she already felt insecure as a parent. At a minimum, users should be able to determine whether their conversation content is used for personalized advertisements.

This study has limitations. First, we did not test rule-based chatbots. However, we believe that the CAPE framework can also be applied to rule-based chatbots without modification. Our evaluation focused on OpenAI’s custom GPT chatbots because they are frequently used, and the GPT Store provides a convenient way to determine chatbot usage. Future work should evaluate psychotherapy chatbots on different platforms. In addition, the CAPE framework is currently based on text data. Further iterations should consider voice and visual interaction with conversational agents, which is increasingly used and may be the only form of interaction for populations with certain disabilities. We tested the framework in English and Spanish to determine multilingual capabilities. OpenAI reports availability in several other languages, which warrants further examination. Finally, although not the focus of this evaluation framework, other features, such as the temporal order of conversation components such as problem exploration before problem-solving, may be important to users or the therapeutic process [28]. Our personalized approach would allow us to assess these and other features.

Second, as AI-based chatbots are increasingly used for clinical purposes, the frameworks to evaluate them must consider ethical concerns. This is especially important in mental health, where ethical violations can negatively affect treatment effectiveness [54,55]. Currently, the CAPE framework only focuses on accuracy and reliability through the lens of therapeutic style and does not penalize chatbots for providing inaccurate or misleading information. Because LLMs may be perceived as trustworthy as humans due to their humanlike conversational abilities [56-58], future iterations of the framework should assess information accuracy. Moreover, users, such as those with social anxiety [59,60], who may be inclined to avoid human therapists, are potentially more vulnerable to misinformation, especially in the absence of external verification. Relatedly, evaluating how a chatbot response is reached is needed. Unlike human therapists, who can explain their reasoning and thereby foster trust, AI chatbots lack this explainability [50,61]. None of the chatbots we analyzed offered a way to check their responses against their training data. This inability to validate chatbot statements increases the risk of misinformation and resulting harm.

Finally, while the CAPE framework addresses the presence or absence of a privacy policy, future iterations should also focus on data storage and protection, with the aim of ensuring the confidentiality of potentially sensitive medical information [50].

In conclusion, the CAPE framework is a promising tool for assessing the quality of psychotherapy conversational agents. We believe this effort will support the development of conversational agents that are safe, accessible, evidence-based, and engaging.