Review Reports - AI in Psychotherapy: Opportunities and Risks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors, the article is very interesting and addresses an important topic; however, I have several suggestions:

Although a review paper does not necessarily have to follow the methodology of a systematic review, it would be useful to briefly describe the process of literature search and source selection in order to better assess the scope, relevance, and balance of the literature presented.

In the manuscript, there is occasional overlap between general-purpose AI systems, AI companionship tools, chatbots, and clinical LLMs, although these are conceptually and functionally different tools.

The risks associated with general AI tools are sometimes implicitly extended to clinically oriented systems without sufficient distinction and nuance.

As a result, the manuscript is not fully balanced in relation to its title and stated aim.

The potential applications of AI, for example in therapist training, supervision, psychoeducation, and support for low-risk interventions, could be presented in a more systematic and analytically developed manner.

This is particularly relevant for the sections discussing emotional dependency, parasocial attachment, and the potential negative effects of AI interactions.

It is not sufficiently clear to what extent certain concepts (e.g., AI psychosis) are theoretically established and how widely accepted they are in the relevant scientific and professional literature.

In the text, it is not always clearly distinguished what represents an empirically supported finding, what is a theoretical interpretation, and what is a normative or speculative claim. A clearer distinction between these levels of argumentation would further strengthen the manuscript.

The recommendations for the development of safe therapeutic AI systems remain rather general, and concepts such as human oversight, safety alignment, and regulatory frameworks should be more concretely connected to psychotherapeutic practice and real-world conditions of implementation.

Practical clinical implications are not sufficiently elaborated. It would be useful to more clearly define the boundaries of acceptable AI use in psychotherapy and to outline possible safety protocols for high-risk situations.

Author Response

Dear Authors, the article is very interesting and addresses an important topic; however, I have several suggestions:

The risks associated with general AI tools are sometimes implicitly extended to clinically oriented systems without sufficient distinction and nuance.

As a result, the manuscript is not fully balanced in relation to its title and stated aim.

This is particularly relevant for the sections discussing emotional dependency, parasocial attachment, and the potential negative effects of AI interactions.

Thank you for taking your time to review my text!

While selecting the literature I chose to review, the main factor has been how recent the papers were published. AI development (and its issues) shift rather rapidly, at a scale of weeks and months, not years, as the models evolve. For example, while I was researching the response conversational AI agents give to high risk messages, the software in question went through multiple versions and even changed the model itself.

I also strived for the papers to be mostly open access in order for others to be able to review them themselves. It’s part of the reason why I chose to publish here, since easily accessing knowledge is becoming more crucial than ever. I have added a methodology section.

The overlap between conversational and clinical LLMs, from a psychotherapy point of view, is due to the nature of therapy itself. Symbolic AI can, in theory, be used in psychotherapy to score tests from patients, but little else, since defining every possible problem and outcome is a sisyphean task. The biggest advantage of this type of tool is being able to retrace the steps in every response with a series of ifs and thens, whereas each patient has such a unique life experience it is currently impossible to use them beyond statistics and clinical tests.

I’ve modified the article text to better highlight this aspect, I appreciate the insight.

Regarding the potential applications of AI in therapy, the recommendations were more general since my field of expertise is psychology, not informatics and computer science, so I can’t confidently comment on specific limitations and possibilities.

Designing a therapeutic AI tool where each response has a traceable logic, peer-reviewed by trained professionals around the clock, that can handle mild day-to-day problems and instill a sense of trust and safety in the user while accurately assessing when healthy boundaries are breached, and not put users in danger all while somehow respecting ethical and legal requirements is simply science-fiction at the moment.

Currently, the best possible use for AI in a therapeutic setting is in therapist training, but even there the risk of misuse remains, since they could, in theory, use real life patient data to generate responses. In my opinion this risk is far less more damaging than unsupervised use of conversational agents as a replacement for therapy. I have made the necessary modifications in order to show this better. I have also modified the phrasing where it was not sufficiently clear.

Regarding the safety protocols, I cannot possibly define something that varies wildly between AI models, and between their versions, since each generative AI tool has its own self made logic that cannot be modified or even viewed. Again, I acknowledge my limitations given the lack of overlap between computer science and my field, but I cannot speculate on something this important when it could simply be impossible to implement given the current state of technology.

Reviewer 2 Report

Comments and Suggestions for Authors

With one major exception, I am favorably disposed toward this review of the risks and opportunities associated with utilizing AI in psychotherapy. I recommend the following revisions.

Any study should clearly state its research problem (the "What?" or objectives of the study) and convey its significance (the "So what?" or unique/essential contributions of the study). The author should add a paragraph to the introduction addressing these considerations. "This study aims to... The knowledge conveyed by this review is important because..." This information should form a new paragraph of 4-6 sentences.
The methodology the author used to generate this review must be explicitly stated in the manuscript for scientific consideration and replicability. Typically, there is a section called Materials and Methods (or, if preferred, Review Methodology) for this purpose. I recognize that journal sections and mandates for review articles vary, but review methodology information is essential. A section like that must be added and should address the following considerations: (a) What type of review is this? It does not seem to be a systematic review or a scoping review, but it is an evidence-based review of prior studies that draw on valid and reliable data? Specify the type of review you conducted while describing its key attributes and underlying logic. (See the journal's encouragement to explain review procedures here: https://www.mdpi.com/journal/behavsci/instructions. (b) What methodological procedures were used to conduct the review? Attend to this point for each phase of the study completion process. Identify the databases searched, the search terms employed, the inclusion/exclusion criteria used to discern suitable publications to review from those deemed unsuitable, and the number of articles identified at each phase of the review process. Authors will commonly say that X number of articles were initially returned by databases but that a smaller Y number of articles made the final sample after selection criteria were applied. If there were theoretical considerations (e.g., attachment) that governed searches and domains of inquiry, be explicit about those as well as empirical areas of focus.
Some review authors generate a table of all the articles in the final sample of publications reviewed with key attributes of those articles listed in the table. I'd encourage this but say it's optional. Consider engaging with review procedures indicated by PRISMA (https://www.prisma-statement.org/), which is the gold standard of review articles and is mentioned in this journal's Instructions for Authors. The author may wish to read through PRISMA procedures and describe which, if any, informed this review. I am not suggesting every facet of PRISMA needs to be followed because this is not a scoping review. However, if another review methodology was used, I'm open to that. And if no particular methodology was used, the author still needs to explicate what was done to produce the sample of reviewable articles and the techniques used to analyze them. In a sense, the articles included in the sample of articles are your "data" and should be treated as such through clearly explicated data collection and analysis procedures. Core point: A study methodology section of at least a few paragraphs addressing these considerations would greatly enhance this study.
The citations are references are sometimes problematic. For example, this link (https://www.sciencedirect.com/science/article/abs/pii/S1566253525004257) directs me to an actual article with authors listed. Why is it embedded as a link rather than a standard citation to a reference? Please check throughout so that authors are listed in standard citation format.
Small points that reveal attention to detail in manuscript preparation: (a) At the first mention of artificial intelligence in the text (lines 28-29), add the AI abbreviation in parentheses immediately after that. I appreciate that abbreviations are specified at the end of the article, but this would still be a welcome addition at first mention for any abbreviation. (b) Ensure consistent and correct capitalization of all headings. I see various forms of capitalization used. (c) Similarly, the citations are typically not separated by semi-colons. Commas featured after both the author's name and between the citations are confusing.

The substance of this manuscript is impressive, but a detailed methodology section is needed so that anyone could replicate your results. Attention to other details listed above will also inspire greater confidence in the results reported here.

Author Response

With one major exception, I am favorably disposed toward this review of the risks and opportunities associated with utilizing AI in psychotherapy. I recommend the following revisions.

Any study should clearly state its research problem (the "What?" or objectives of the study) and convey its significance (the "So what?" or unique/essential contributions of the study). The author should add a paragraph to the introduction addressing these considerations. "This study aims to... The knowledge conveyed by this review is important because..." This information should form a new paragraph of 4-6 sentences.
The methodology the author used to generate this review must be explicitly stated in the manuscript for scientific consideration and replicability. Typically, there is a section called Materials and Methods (or, if preferred, Review Methodology) for this purpose. I recognize that journal sections and mandates for review articles vary, but review methodology information is essential. A section like that must be added and should address the following considerations: (a) What type of review is this? It does not seem to be a systematic review or a scoping review, but it is an evidence-based review of prior studies that draw on valid and reliable data? Specify the type of review you conducted while describing its key attributes and underlying logic. (See the journal's encouragement to explain review procedures here: https://www.mdpi.com/journal/behavsci/instructions. (b) What methodological procedures were used to conduct the review? Attend to this point for each phase of the study completion process. Identify the databases searched, the search terms employed, the inclusion/exclusion criteria used to discern suitable publications to review from those deemed unsuitable, and the number of articles identified at each phase of the review process. Authors will commonly say that X number of articles were initially returned by databases but that a smaller Y number of articles made the final sample after selection criteria were applied. If there were theoretical considerations (e.g., attachment) that governed searches and domains of inquiry, be explicit about those as well as empirical areas of focus.
Some review authors generate a table of all the articles in the final sample of publications reviewed with key attributes of those articles listed in the table. I'd encourage this but say it's optional. Consider engaging with review procedures indicated by PRISMA (https://www.prisma-statement.org/), which is the gold standard of review articles and is mentioned in this journal's Instructions for Authors. The author may wish to read through PRISMA procedures and describe which, if any, informed this review. I am not suggesting every facet of PRISMA needs to be followed because this is not a scoping review. However, if another review methodology was used, I'm open to that. And if no particular methodology was used, the author still needs to explicate what was done to produce the sample of reviewable articles and the techniques used to analyze them. In a sense, the articles included in the sample of articles are your "data" and should be treated as such through clearly explicated data collection and analysis procedures. Core point: A study methodology section of at least a few paragraphs addressing these considerations would greatly enhance this study.
The citations are references are sometimes problematic. For example, this link (https://www.sciencedirect.com/science/article/abs/pii/S1566253525004257) directs me to an actual article with authors listed. Why is it embedded as a link rather than a standard citation to a reference? Please check throughout so that authors are listed in standard citation format.
Small points that reveal attention to detail in manuscript preparation: (a) At the first mention of artificial intelligence in the text (lines 28-29), add the AI abbreviation in parentheses immediately after that. I appreciate that abbreviations are specified at the end of the article, but this would still be a welcome addition at first mention for any abbreviation. (b) Ensure consistent and correct capitalization of all headings. I see various forms of capitalization used. (c) Similarly, the citations are typically not separated by semi-colons. Commas featured after both the author's name and between the citations are confusing.

Thank you for your time to review my text, I appreciate it!

The articles were chosen based on how recent and relevant the information was, given the fast paced nature of the subject at hand. I strived for the articles and studies to be open access as much as possible, sadly not all of them are. Ease of access to information is something of great importance for me in my research.

I have added a brief methodology section at the end of the article in order to cover this aspect, thank you for taking the time to comment on this. Sadly it would be rather difficult to retrace my steps in the initial research phase since it spanned several months, trickled between courses, clients and supervisions, so I cannot accurately say how many articles I read and considered before settling on the current ones. I could offer an estimate but I feel like that defeats the purpose since it would instill more uncertainty. I will keep this in mind moving forward.

I have also modified the introduction, given the feedback you offered, to showcase the scope of the review more concisely.

The citations and capitalization errors were corrected.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Author,

Thank you for the clarifications and the incorporated corrections. I have no further suggestions, and my recommendation is to accept the article.

Author Response

Thank you for the clarifications and the incorporated corrections. I have no further suggestions, and my recommendation is to accept the article.

Thank you.

Reviewer 2 Report

Comments and Suggestions for Authors

Well done on this revision. State that your lack of a formal replicable methodology is a limitation in the methods section. Please also add the problem statement and significance paragraph to the Introduction in 4-6 sentences as initially requested. The abstract was revised but the intro still needs that paragraph. Otherwise, well done.

Author Response

State that your lack of a formal replicable methodology is a limitation in the methods section. Please also add the problem statement and significance paragraph to the Introduction in 4-6 sentences as initially requested. The abstract was revised but the intro still needs that paragraph.

Thank you again for taking the time to reply! I have added the neccessary paragraph in the introduction, and also the disclaimer in the methodology section.