The rise of artificial intelligence for cognitive behavioral therapy: A bibliometric overview

Abstract

Recent years have seen a sharply rising interest in the scientific area dedicated to the study of the use of Artificial Intelligence (AI) for Cognitive Behavioral Therapy (CBT) research and applications (AI4CBT for brevity). Yet, little is known about how this interest is realized and hence the overall status, prospects, and possible challenges of AI4CBT as a field (e.g. breadth of the field, key topics and methods, key producing countries/institutions/authors, interdisciplinary grounding). This paper addresses this gap by developing a broad‐spectrum bibliometric analysis towards acquiring a comprehensive overview of the AI4CBT field. Four key dimensions are analyzed (productivity, producers, productions, and contents) along the array of bibliographic metrics, including production trends over time, leading contributors at various levels, co‐authorship, citation, and keywords co‐occurrence networks, publication formats, key venues, methodological trends, and disciplinary assessment. The paper concludes by framing the status of AI4CBT as a scientific field, allowing to tie it to scientific and applicative challenges and opportunities that AI4CBT may encounter and offer as it further develops.

INTRODUCTION

Cognitive Behavioral Therapy (CBT) is a key area of interest in both research and society, as it offers an accessible, evidence‐based approach to addressing a wide range of mental health issues, including anxiety, depression, and stress (Furukawa et al., 2021; Hofmann et al., 2012; Karyotaki et al., 2021; Nakao et al., 2021; Rothbaum et al., 2000; Soh et al., 2020; Van Dis et al., 2020). CBT is a psychotherapeutic approach grounded in the assumptions that psychological problems are based on, at least partly, faulty or unhelpful thinking patterns as well as on learned unhelpful behavior patterns and that these problems, symptoms, and harm to individual functioning can be alleviated through learning new ways of coping with them (Association AP, 2017). CBT treatment usually focuses on changing thinking patterns (e.g., recognizing and addressing deliberative distortions, better understanding behavior and motivation of others, improving problem‐solving abilities and self‐confidence); as well as behavioral patterns (e.g., facing instead of avoiding fears, roleplaying potentially problematic interactions, learning to calm one's mind and relax one's body). In these therapies, the therapist primarily supports the patient in scaffolding a learning process, usually through in‐session exercises and out‐of‐session homeworks, focusing on improving the functioning of the patient rather than investigating the source of the difficulties (Association AP, 2017).

Although not universally superior to other psychotherapies or pharmacological treatments (Cuijpers et al., 2013), CBT is one of the most widespread approaches for promoting mental health awareness and resilience (David et al., 2018; Happer et al., 2017; Sánchez‐Ortiz et al., 2011) through a broad variety in its implementation methods, aims, and contexts. CBT can be delivered through multiple media, from in‐person support to digitalized support (Bee et al., 2010; Forman & Barakat, 2011; Kojima et al., 2010; Wan Mohd Yunus et al., 2022) and Internet‐Based CBT (Andersson, 2024a; Sánchez‐Ortiz et al., 2011). CBT has accrued an extensive empirical demonstration of its relevance for an array of therapeutic aims, including depression (Andersson, 2024b), bulimia (Sánchez‐Ortiz et al., 2011), insomnia (Trauer et al., 2015), anxiety disorders (Bandelow et al., 2015), chronic fatigue syndrome (Malouff et al., 2008), obsessive‐compulsive disorders (Olatunji et al., 2013), post‐traumatic stress disorder (Shubina, 2015), panic disorders (Papola et al., 2023), psychosis (Hutton & Taylor, 2014), schizophrenia (Turkington et al., 2004), stress (Nakao et al., 2021), smoking cessation (Gierisch et al., 2012), chronic pain (Knoerl et al., 2016). CBT has been applied in a variety of settings, from individuals in a clinic to schools and workplaces, and within a variety of healthcare systems (Joyce et al., 2016; Kambara & Kira, 2021). CBT is often argued to be accessible to a broad variety of users, through a flexibility in its delivery methods for example, patients living in remote or geographical locations or with a shortage of trained professionals, coverage of individuals with lower income, through time‐bounded plans, lowered financial costs, and increased cost‐effectiveness (Ali et al., 2024; Hollinghurst et al., 2014), availability to individuals with resistance to medication (Hollinghurst et al., 2014). From a systemic standpoint, CBT, through a patient‐centric approach that may be less demanding on trained professional time, is argued to increase the accessibility of mental health by supporting the scaling up of the availability of professional support, which faces a major shortage (Ahern et al., 2018; Cowain, 2001; Espie & Henry, 2023; Gorenstein & Papp, 2007; McClure et al., 2024; O'Neill & Feusner, 2015), such as serving millions of patients in the UK through the Improving Access to Psychological Therapies initiative (Wakefield et al., 2021). While challenges remain, including treatment adherence and ensuring effectiveness across diverse populations, CBT's structured, evidence‐based design and demonstrated cost‐effectiveness make it a uniquely accessible therapeutic option for individuals and healthcare systems alike.

Artificial Intelligence (AI) methods present significant potential for advancing both CBT research and its societal impact, and this potential is increasingly being realized in practice (Atzil‐Slonim et al., 2024). AI is an umbrella concept covering an array of research interests and operational methods centered on making decisions using artificial substrates (commonly computers). AI research and methods usually emphasize e.g. finding optimal solutions for problems, identifying patterns in data, classifying entities based on their characteristics, and replicating human‐like behaviors. ¹ In research, AI solutions can enable the exploration of new research questions, enhance existing methodologies, and enhance researchers' productivity. For instance, AI can facilitate large‐scale studies by automating the processing of both quantitative and qualitative data (Ewbank et al., 2020; Varadarajan et al., 2023), tapping into new data sources such as social media and biometric information (Clark et al., 2014; Davies, 2023; Laugharne et al., 2023; Straube et al., 2014), and offering advanced analytical methods like topic modeling and automated classification (Prasad et al., 2023; Schmidt et al., 2023; Sundermann et al., 2017; Tao et al., 2023). AI can also support innovative experimental designs (e.g., biometrics, adaptive support systems) and automate tasks such as chatbot‐driven pre‐screening of research participants (Leo et al., 2022), making previously complex research processes more feasible and opening up new possibilities for investigation. In society, AI can also enhance the accessibility and delivery of CBT, as well as its overall impact. Many previously identified activities, such as patient intake (Sikström et al., 2023; Varadarajan et al., 2024), follow‐up, journaling (Shidara et al., 2024; Wiegersma et al., 2020), and the prediction of dropout or treatment outcomes (Duhne et al., 2022; Giesemann et al., 2023; Gómez Penedo et al., 2023; Poster et al., 2021; Prasad et al., 2023), can benefit from automation and data‐driven insights. AI can also assist in CBT training (Parsons et al., 2023) and support institutions and professionals (e.g., optimizing patient allocation) (Delgadillo et al., 2020). Moreover, AI enables innovative formats for delivering CBT, such as guided sessions via chatbots, apps, virtual reality (personalized experiences, virtual characters), and virtual assistants, offering users immediate, round‐the‐clock support without the need for a human therapist (Davies, 2023; Fitzpatrick et al., 2017; Woo et al., 2023; Woo et al., 2024). These systems can track users' emotional patterns, provide personalized exercises, and deliver real‐time feedback based on CBT principles (Miyamoto et al., 2023). While creating promises for benefits, the use of AI technologies for CBT also embeds potential risks that demand dedicated scrutiny. These risks may arise from AI technologies (e.g. algorithmic bias, unintended leakage of private and sensitive information, black boxes) or their use (e.g. de‐humanization of therapies, overreliance, uberization, deskilling) (Lindgren, 2023a; Mulukuntla, 2022; Panch et al., 2019; Sucala et al., 2017; Wasil et al., 2021).

However, despite the significant potential (and risks) of the use of AI methods within CBT research and applications and the growing evidence that this activity is growing rapidly, this activity has not yet been framed as a scientific field. To enable such a framing and subsequent structured scrutiny, this paper proposes to coin the (now‐unlabeled) array of activities relative to the use of AI methods within CBT research and applications (or, more concisely, AI4CBT) as a scientific field. As a field, it remains unclear whether AI4CBT is a dormant prospect, an emerging research topic, a well‐established field with defined structures, or a fading area of interest, beyond general assumptions derived from insights into AI and mental health (Jiang et al., 2024). Given the highly interdisciplinary nature of AI4CBT, ² along with the challenges of conducting it –such as accessing the necessary skills, participants, technologies, authorizations, logistical resources, and institutional support –it is crucial to understand the current state of the field. This understanding is essential for organizing, enabling, and promoting the production of high‐quality AI4CBT research. Furthermore, considering the societal implications (for clinicians, patients, and social services) of integrating AI into CBT, gaining an early overview of the field is essential. In contexts where the adoption of AI technologies has sometimes had negative impacts on affected communities (Lindgren, 2023a; Lindgren, 2023b; Lindgren, 2023c), ensuring that AI4CBT is thoughtfully incorporated into technologies and therapies is key to ensuring it serves the best interests of mental health and the broader social good.

In response, this paper aims to provide an initial answer to the general question of characterizing the status of AI4CBT as a research field by developing a broad overview across several dimensions: What is the size of the field? Is it expanding? Is it gaining interest? From national, institutional, and individual perspectives, who is producing AI4CBT research? Is it organized into a dedicated community building a cohesive body of work? Where, in what format, and for whom is this research being published? Lastly, does the field uphold the interdisciplinary promise of integrating AI and CBT?

This paper addresses these questions through a bibliometric analysis of the AI4CBT field based on the methodology from (Öztürk et al., 2024), for developing an evidence‐based broad‐spectrum overview of how the field is structured. This overview is built over the integration of four key dimensions that can be assessed through specialized bibliometric analyses, themselves quantified from bibliometric data. These dimensions are the productivity of the field (in Section 4.1); the producers at the national, institutional, and individual levels (in Section 4.2); the productions of the AI4CBT field (in Section 4.3), and the contents of the field (in Section 4.4). The specific bibliometric analyses rely on standard bibliographic measures outlined by (Mukherjee et al., 2022) supplemented by specialized bibliometric measures of interdisciplinarity. These measures include: publication metrics, trends in scientific production over time, identification of key producers at national, institutional, and individual levels, co‐authorship networks, publication formats, disciplines, venues, citation networks, keyword co‐occurrence networks, the frequency of technological terms, and disciplinary citation analysis. The insights of these four dimensions are integrated into the discussion (in Section 5) to provide a comprehensive overview of the status of AI4CBT as a field as well as the common challenges and opportunities usually tied to such a status.

RELATED RESEARCH

To our knowledge, no related work provides an overview of the AI4CBT as a scientific field as done in this paper. The most related research consists of surveys and literature reviews. The closest review addresses how AI has been used along the various phases involved by CBTs (i.e. pre‐treatment, therapeutic process, and post‐treatment) (Jiang et al., 2024). Other related reviews focus on AI and mental health more generally, sometimes mentioning CBT as one method among many. (Aggarwal et al., 2023; Boucher et al., 2021; Li et al., 2023) focus on chatbots for promoting mental health behavioral changes and concurred that CBT is the most often used therapeutic background for chatbots, though not informing on how chatbots relate to the overall CBT literature. (Higgins et al., 2023) focuses on AI‐powered decision support systems in mental health, retrieving that only one study mentions CBT.

This related research differs from the current paper in several ways. First, related research relies on surveys and literature reviews, hence primarily providing qualitative insights and relying on a lower number of papers (N<100), sometimes without a structured method for data collection. In contrast, the current paper relies on structured data collection methods for acquiring quantitative insights on the global dynamics of AI4CBT as a research field (e.g., the number of papers per year, key disciplines involved) for a large number of papers (N=935). Second, only one study.

(Jiang et al., 2024) focuses on connecting AI and CBT and restricts its focus on how AI can be used within CBT therapy –hence disregarding other uses of AI for CBT, such as AI for CBT research (e.g., machine learning for retrieving patterns in CBT‐related data). Comparatively, the current paper considers and seeks to elicit all possible perspectives on AI4CBT research. In other papers, CBT is only regarded as one method among many and, besides eliciting a few examples, do not provide any insights on how the CBT field is structured.

Another line of research involves bibliometric analyses that tangentially touch upon AI and CBT. (Chen et al., 2024; Rajkishan et al., 2024) study AI for mental health in student populations, with key results, each only evoking one AI4CBT paper as a passing remark.

(Zale et al., 2021) study technology in psychological interventions and list one paper on computerized CBT. (Soares et al., 2020) focuses on trends in psychotherapy, referring to CBT as the subfield having received the most attention over the last 50 years and mentioning a growing interest for AI in psychotherapy as a passing remark, with no explicit connection to CBT. (Wang et al., 2024) study psychological interventions for depression in children and adolescents and mention the prospective relevance of AI as a passing remark. While some of these studies remotely connect to AI4CBT through passing remarks, none of them provide any quantitative insights on AI4CBT as a field, which is the aim of the current paper.

METHOD

Data collection and filtering

Bibliometric analyses are a common research method (Donthu et al., 2021) for identifying statistical patterns from the bibliographic metadata of large corpora of scientific documents (N>100). The standard process for conducting a bibliometric analysis involves the following steps: 1) selecting a database, 2) acquiring documents from this database, typically via a search query, 3) retrieving the metadata of documents linked to this query, 4) filtering out irrelevant documents, 5) retrieving the metadata of entities related to the filtered documents (e.g., authors, institutions), 6) developing metrics to gather quantitative evidence to address the research questions, and 7) analyzing this evidence to generate insights relevant to the research questions. A summary of the key steps in the data collection process is provided in Figure 1.

FIGURE 1.

Data collection and processing flowchart.

Database

The Scopus database, accessed through the Scopus API, was selected due to its ability to provide structured data with unambiguous unique identifiers for each document, as well as options for advanced algorithmic database queries and statistical manipulations (e.g., retrieving all documents cited by or citing the documents within the corpus). Additionally, Scopus is one of the largest available databases and is routinely used for conducting bibliometric analyses (Schotten et al., 2017; Stahlschmidt & Stephen, 2020).

Search query

The following query was used to retrieve the document IDs tied to AI4CBT (the query is exhaustively introduced in Annex A).

$$\text{Query}=\mathrm{AI}\text{AND}\mathrm{CBT}$$

where AI and CBT are both defined based on a set of keywords, as a sequence of OR statements. For each keyword k from each set, the query encompasses any document that contains k in its title, abstract, author‐defined keywords, or publication venue. For example; the AI query is defined asTITLE(“artificial intelligence”) OR TITLE(“machine learning”) OR … OR ABS(“artificial intelligence”) OR ABS(“machine learning”) OR …

CBT keywords

The set of keywords for CBT was constructed by compiling the keywords of the CBT‐related component of the queries from (Soares et al., 2020; Xin et al., 2022). This set is: "cognit* behav* treat*", "cognit* behav* therap*", "cognit* behav* psychotherap*", "cognit* behav* interv*" and covers all the plausible spellings for CBT. The acronym CBT was left out due to introducing numerous false positives (e.g., Computer‐Based Teaching). For the sake of ensuring congruence with former studies and for avoiding the introduction of biases (towards e.g., a specific framing of CBT or CBT de‐livery method) in a context where the conceptual boundaries of CBT and the methods for delivering CBT remain the object of academic debates (Baardseth et al., 2013; Fang & Ding, 2023; Zlomke & Davis, 2008), this study left out related therapeutic concepts (e.g., "exposure therapy", "acceptance and commitment therapy") and due to the possible introduction of false positive from other fields (e.g. "third wave approach"). Naturally, the addition of these terms could help broaden future research avenues and obtain a more specific mapping of the many shapes CBT can take.

AI keywords

The set of keywords used to capture the rapidly evolving terminology associated with AI was constructed by compiling keywords from several sources. These sources include: 1) the keywords used to describe AI in related works on AI and mental health (Chen et al., 2024; Rajkishan et al., 2024); 2) AI‐specific keywords identified in the table of contents of one of the most widely used AI textbooks (Rani et al., 2015); 3) the AI glossary developed by Science (Glossary, 2017); 4) recent emerging trends in AI (e.g., ChatGPT); 5) related keyword‐based bibliometric studies involving AI (Vanhée & Borit, 2023); 6) author‐keyword‐based snowball sampling; and 7) semantic‐based snowball sampling. For steps 6 and 7, a snowball sampling approach was applied. In this method, an initial corpus is collected, and a selection criterion is used to identify new keywords that may be relevant for detecting AI methods. If a new keyword is identified, it is added to the list, and the process is repeated with the updated corpus. If no new keyword is found, the process halts. For step 6, the identification method involved searching for AI‐related keywords among all author keywords that appeared at least 10 times. For step 7, the identification method involved sorting all documents that cited or were cited by the corpus by their semantic proximity to the average semantic value of the corpus. This sorting was performed as follows. First, a semantic vector (also known as an embedding) was computed for each document to represent its semantic content, using a BERT encoder applied to the document's title, abstract, and keywords (Koroteev, 2021). ³ Second, the average semantic vector for the corpus was calculated by averaging the semantic vectors of all the documents in the corpus, similar to the method in Reimers (2019). Third, for each document that cited or was cited by the corpus, the semantic vector was computed and compared to the average semantic vector of the corpus to determine its semantic closeness. Fourth, citations that did not reference CBT were removed, and the remaining documents were sorted by their semantic distance from the average semantic value of the corpus. Fifth, the title, abstract, and keywords of the top 20 documents were reviewed manually to assess the presence of relevant AI‐related keywords. Neither step 6 nor step 7 resulted in the identification of new keywords. The most relevant citations pertained to the digitalization of CBT, without specifying AI‐related terms. Finally, ambiguous keywords, such as “reasoning under uncertainty", and acronyms were removed to avoid generating false positives. Although necessarily incomplete, it is reasonable to assume that, unless the AI4CBT field consists of entirely disconnected communities, the selected keywords are sufficiently comprehensive, and the collected corpus is representative of the AI4CBT research field. The query retrieved 946 documents. The cut off date of data collection was 26.09.2024.

Eliminating false positives

Two steps were implemented to assess the presence of false positives. First, a random manual assessment was conducted on N=30 documents, along with manual checks during the setup of the data processing pipeline. Second, the documents in the corpus were sorted by their semantic proximity to the average semantic value of the corpus, using the same BERT‐based semantic distance metric applied in step 8 of the data collection method. The top 20 documents that were furthest from the average were analyzed manually for relevance, and only two were identified as false positives. This process complements the elimination of ambiguous keywords and acronyms.

As a bibliometric note, the current query captures any document intersecting AI and CBT, which naturally includes AI4CBT documents but could also, in principle, include other combinations, such as studies on CBT for supporting stressed AI work environments. To ensure that the documents in the corpus are indeed focused on applying AI methods to CBT research, a thorough manual assessment was conducted on 100 randomly selected documents, plus the 20 most divergent documents identified through the false‐positive elimination method. All documents were confirmed as involving AI methods for CBT research and application, providing strong confidence that the corpus is highly representative of the AI4CBT field.

Additionally, the number of documents retrieved for each year for both the AI query and the CBT query was recorded. This step involved performing a series of queries (e.g., AI AND PUBYEAR 2024 for the number of publications related to AI in 2024) for each year in which at least one AI4CBT publication was identified.

Filtering

Documents were filtered out of the corpus as follows. Retracted documents were removed (N=1), leaving N=945 remaining documents. Conference reviews were excluded (N=10) as they do not contain substantive content (they only list the documents accepted in a conference without offering relevant metadata). After this, N=935 documents remained (i.e., 99% of the original query).

Data processing

The data was extracted via the Scopus API using a custom Java program. This step ensured that all documents, authors, and institutions were uniquely identified and prepared for further processing. For each document, basic metadata was retrieved (including title, author IDs, author affiliations as institution IDs, subject area codes, etc.), as well as the set of document IDs citing and cited by the documents in the corpus. Basic information for these citations was also collected. For each retrieved author ID, the author's name was retrieved, and for each institution ID, the institution's name and country were obtained.

Data analysis: attributing disciplines to documents

The Scopus database is frequently used for conducting disciplinary analyses. This database assigns a set of subject areas to each document, based on the All Science Journal Classification (ASJC) framework. ⁴ These subject areas provide key information about a document's discipline. The ASJC framework comprises 334 categories, each with a unique identifier and title. Each ASJC category belongs to one of 27 Subject Area Classifications, which represent broad scientific fields. For example, the ASJC code 3103, titled “Astronomy and Astrophysics" belongs to the “Physics and Astronomy" Subject Area Classification. A set of subject areas is manually assigned by Scopus reviewers to each venue, and indirectly to each document. A document may be assigned multiple ASJC codes. These codes are ranked based on the relative CiteScore metric, which measures the scientific impact of a venue by the average number of citations per document over recent years, compared to other venues within the same ASJC category ⁵ (Fang, 2021). In this paper, we refer to the primary code as the ASJC code listed first among the codes attributed to a document.

To simplify the analysis, we consolidated the 27 Subject Areas from the original ASJC framework into nine broader groups, which we refer to as “disciplines”, following a similar approach to (Vanhée & Borit, 2024). The mapping of these disciplines is detailed in Table 1. To ensure the accuracy of this classification, 20 blind assessments of randomly selected documents were performed. In all trials, the primary discipline was correctly identified except in cases involving the overlap between the mind and medical sciences, which is likely due to the close relationship between these two categories in the AI4CBT context. Therefore, special attention was given to processing these two disciplines. In our approach, we treated them as equivalent, with a note indicating if they produced significantly different results (which they did not).

TABLE 1.

Overview of the disciplines used in this study. Note that the subject area classification for psychology is under the mind sciences discipline. The social sciences subject area classification refers to society‐related sciences (e.g. archaeology, business and economics, sociology). CBT, through its mental health component, conceptually overlaps both medical sciences (which lists e.g. mental health under medicine and nursing, and psychiatry under medicine), and mind sciences (which lists clinical psychology under psychology).

Disciplines	Subject areas classifications
Arts and humanities	Arts and humanities
Computational sciences	Computer science; engineering; mathematics
Economic sciences	Decision sciences, business; economics, econometrics and finance
Environmental sciences	Agricultural and biological sciences; energy; environmental science; earth and planetary sciences; veterinary
Multidisciplinary sciences	Multidisciplinary
Medical sciences	Biochemistry, genetics and molecular biology; immunology and microbiology; medicine; nursing; pharmacology, toxicology and pharmaceutics; dentistry; health professions
Mind sciences	Neuroscience; psychology
Physical sciences	Chemistry; chemical engineering; materials science; physics and astronomy
Society sciences	Social sciences

RESULTS AND DISCUSSION

AI4CBT productivity

Field‐scale, individual productivity, turnover: key bibliometric figures

Table 2 summarizes the key variables related to the AI4CBT corpus, as typically identified in bibliometric analyses. The term “Unique authors/affiliated institutions/countries" refers to the total number of distinct authors, affiliated institutions, or countries. In contrast, “Author/affiliated institutions/countries appearances" refers to the total number of times authors, affiliated institutions, or countries appear in the corpus. For example, if the corpus includes author A who has written three documents, there is one unique author and three author appearances.

TABLE 2.

Key bibliometric figures of the AI4CBT corpus.

Category	Number
Number of documents in the corpus	935
Number of unique authors in the corpus	4,616
Number of author appearances in the corpus	5,427
Number of unique affiliated institutions in the corpus	1,671
Number of affiliatations reported in the documents of the corpus	6,463
Number of countries in the corpus	70
Number of unique documents cited‐in by the corpus	34,498
Number of unique documents citing out the corpus	15,113
Average contribution per author (average number of documents per author divided By the number of co‐author per document)	0.20
Number of appearances per author	1.18
Number of co‐authors per document	5.80

The scale of the field can be inferred from the number of documents, authors, and institutions. The numbers in the table suggest that the scale is in the thousands of documents and authors, indicating that AI4CBT is an intermediate‐sized field. While it has garnered significant converging attention, it remains far from being a fully consolidated research field, which would typically involve tens of thousands of publications.

The degree of individual productivity, individual stability/turnover, and document‐level collaborative effort can be inferred from the values at the bottom of Table 2. A low individual contribution per author is observed, as each author produces, on average, N=0.20 documents while appearing in N=1.18 documents. These values suggest a high turnover, where most authors contribute to only a single document before leaving the field. The number of co‐authors per document is relatively high (n=5.80), indicating that AI4CBT publications typically involve a large number of collaborators. This is understandable given the interdisciplinary nature of AI4CBT research and the common co‐authorship practices in psychology and medical research, which are central themes within the AI4CBT domain.

Temporal trends: scientific production over time

Figure 2 represents the number of AI4CBT documents published each year and the proportion of AI4CBT documents relative to the total number of AI documents and CBT documents per year.

FIGURE 2.

Number of publications per year. The black line relates to the left Y axis and represents the number of AI4CBT publications in absolute given the year set by the X axis. The green line (respectively orange line) represents the proportion of the number of publications of AI4CBT relative to the number of publications on AI (respectively CBT). The green and orange lines are scaled such that their maximum is represented by 1 on the Y‐axis on the right side. The value of 1 is associated with the maximum value indicated in the legend, i.e. 0.03% of all AI publications for the green curve and 4.9% of all CBT publications in the orange curve.

This representation allows assessment of the growth of AI4CBT over time, showing a steady rise in productivity since its inception and a sharp acceleration of yearly output starting in 2017. Such a trajectory is typical of emerging fields, which often follow an S‐shaped growth curve as the field matures and interest stabilizes (Mansfield, 1961; Price, 1963; Rogers et al., 2014). The post‐2017 surge in interest may be attributed to the introduction of new technically‐accessible AI techniques that directly apply for AI4CBT research, such as the quantization of textual semantic value through BERT methods (Devlin, 2018) and streamlined, user‐friendly machine learning toolkits. A similar trajectory can be observed in e.g. AI for medicine research (Guo et al., 2020), albeit with the state of the explosion of the field being initiated slightly earlier (2015 instead of 2018 for AI4CBT).

Figure 2 also illustrates the temporal evolution of AI4CBT production relative to its “parent" fields, i.e. AI and CBT. The overall upward trends sharply suggest that AI4CBT has grown rapidly in importance relative to both AI and CBT, especially since 2017. This finding highlights that AI4CBT may be a field attracting a dedicated interest for its development. The growth in AI4CBT's relative importance to CBT is particularly striking, with a fivefold increase since 2017, now reaching an impressive 4.9% of involvement of AI in all CBT publications. This finding highlights that AI increasingly becomes a key component of CBT research practices.

Producers of AI4CBT research

The relatively recent emergence of AI4CBT as a field may make it difficult to identify where related research is being produced. Identifying these research centers is important for understanding the global AI4CBT research landscape (e.g., AI4CBT‐friendly countries, centers of expertise) and pinpointing areas that are conducive to the development of AI4CBT research (e.g., where to find the necessary expertise and technological infrastructure). This section highlights the primary centers of AI4CBT research production at the national, institutional, and individual levels.

National Interests: country‐level production

Figure 3 illustrates the country‐level interest in AI4CBT research by showing the number of documents produced by each country. These figures reveal the typical correlation between high academic productivity and countries with high‐income and/or large populations, as well as a greater interest in CBT among high‐income countries (Kuo, 2019). Notably, the overwhelming presence of the US can be observed, producing nearly three times as many documents as the second most productive country, suggesting a strong national interest in AI4CBT. Conversely, the absence of high‐income Latin American countries in the top ranks can be observed, possibly due to lower interest in CBT, stemming from cultural barriers and other psychological traditions (Bernardelli et al., 2022; Botbol & Gourbil, 2018; Neufeld et al., 2021). This observation is confirmed when considering AI for medicine research (Guo et al., 2020), in which top‐producing countries follow a similar order as AI4CBT except with the inclusion of two Latin countries.

FIGURE 3.

Histogram of the top 10 most productive countries (left) and institutions (right) in the AI4CBT corpus. The number represents the total number of documents in the corpus that list the country/institution within the affiliations of its authors.

Production and geographical distribution of key institutions: institution‐level production

Figure 3 presents the most productive institutions, showing the top 10 institutions that produced the most documents in the AI4CBT corpus. The histogram reveals two distinct groups: the first consists of three institutions with the highest productivity, producing 25–30 documents each; the second consists of the remaining institutions, each producing 10–15 documents. This distribution suggests a possible centralization of AI4CBT research around a few institutions, which is plausible given the field's highly specialized and interdisciplinary nature.

Figure 3 also highlights the geographical distribution of key AI4CBT‐producing institutions. When compared to the country‐wide distribution in Figure 3, it shows a general correlation between top‐producing countries and local top‐producing institutions. A notable exception is Sweden, which stands out with three of the 10 most productive institutions, including the top producer, despite being ranked seventh overall in country‐level production. This suggests a concentration of AI4CBT research within Swedish institutions.

Research Spearhead and Social Factors: author‐level production

Emerging fields are often influenced by various social factors that can significantly shape the field's dynamics, especially in smaller, nascent fields where individual contributions may have an outsized impact. For instance, a field may revolve around a single‐star researcher, making its development highly dependent on that individual; alternatively, there may be no clear leader, suggesting that the field may consist of independently occurring, sporadic research interests (Mintzberg, 1989).

Figure 4 examines the degree of concentration within the AI4CBT research cohort, showing the number of documents authored by the 20 most prolific researchers. The relatively flat distribution indicates that, while there are some disparities, the number of documents per author remains relatively consistent, ranging from 12 appearances for the most prolific author to 5 for the 20th most prolific.

FIGURE 4.

Histogram of the top 20 most prolific authors in the AI4CBT corpus. The number represents how many times each author appears in the corpus.

This observation, combined with the low number of appearances per author noted in Section 4.1.1, suggests that AI4CBT research is driven by a relatively broad spearhead of researchers. Moreover, Figure 4 shows that the most prolific authors are geographically dispersed, though this distribution largely aligns with country‐level productivity. Altogether, this analysis suggests that the field has a relatively stable trajectory over time.

Topology of the research community: co‐authorship analysis

The topology of the research community was assessed through a co‐authorship analysis, which reveals the extent of collaboration within the field. From a bird's‐eye view, such an analysis can illuminate key factors like the closeness or diffuseness of the community and the presence of subgroups. Figure 5 shows the co‐authorship network for all authors with at least three publications (N=123), ensuring a manageable visualization.

FIGURE 5.

Co‐authorship analysis of all authors with at least three publications (N = 123). Each node represents an author, and links between nodes represent co‐authorship. Thicker nodes indicate more publications by the author, and thicker edges represent more co‐authored documents. Colors represent clusters, or groups of authors who frequently collaborate. Closer nodes represent authors with fewer degrees of separation in co‐authorship. The analysis was conducted using VOSViewer.

The figure, composed mainly of small clusters of up to 15 researchers, reveals that the AI4CBT community remains highly fragmented, with little to no connections between groups. This suggests that AI4CBT research is likely being carried out by independent teams with minimal collaboration, hinting at possible duplicate efforts (e.g., developing expertise, software, methods) due to parallel activities in isolation. Given the nascent nature of the AI4CBT field, this fragmentation is understandable but underscores the need for further consolidation of research networks to foster the field's development.

AI4CBT publications

Publication Format & Velocity: key bibliometric figures

Table 3 presents the number of documents per publication format (e.g., conference papers, journal articles, books) related to AI4CBT research, excluding conference reviews, which were filtered out. The data reveals a strong dominance of journal articles, followed by conference papers, book series, book chapters, and books. The fact that journal articles account for 83% of the corpus suggests that AI4CBT research tends to favor a slower‐paced publication of more consolidated results. This trend aligns with the publication practices typically seen in psychology and medical sciences, rather than those of computer science.

TABLE 3.

Number of documents per publication type.

Document type	Number
Conference paper	89
Journal article	773
Edited volumes	0
Research review	0
Popular science presentations	0
Book chapters	19
Book	3
Editorial	0
Conference review	N/A
Book series	51
Trade publication	0
Dataset	0
Report	0
Thesis	0

Disciplinary embedding: publication disciplines

Figure 6 shows, for each discipline, the number and proportion of documents in the corpus classified with that discipline as their primary field.

FIGURE 6.

Number and proportion of primary disciplines for documents in the corpus.

The results, as presented in Figure 6, reveal a strong dominance of three primary disciplines: Medical Sciences (48% of the corpus), Mind Sciences (26%), and Computational Sciences (18%). Other disciplines are marginally represented (<4%). This highlights that AI4CBT research spans its core thematic disciplines, where AI methods are rooted in computational sciences and applied to medical and psychological research. The balance between these three disciplines is noteworthy, as in more established interdisciplinary fields, computational science production often overshadows the domains it supports (Guo et al., 2020; Vanhée & Borit, 2024; Wahle et al., 2023). Thus, the findings suggest that, unlike many other disciplines, AI4CBT is primarily driven by venues focusing on the applied aspects (i.e., CBT) while being partially examined from a methodological standpoint (i.e., AI).

Moreover, it is important to consider the underrepresented disciplines. Given the critical societal and methodological issues related to both AI and CBT (e.g., AI bias, data biases, normativity), one might expect more publications from social sciences (e.g., sociology). This indicates a potential knowledge gap regarding the societal implications of AI4CBT research and applications, which is essential to address as the field is seemingly rapidly expanding into society.

Publication venues

The venues that have published the highest number of AI4CBT documents include: BMC Psychiatry (N=11), BMJ Open (N=11), Journal of Affective Disorders (N=11), JMIR Formative Research (N=12), Journal of Consulting and Clinical Psychology (N=12), PLoS ONE (N=12), Psychotherapy Research (N=14), Journal of Medical Internet Research (N=15), Frontiers in Psychiatry (N=17), and Lecture Notes in Computer Science, including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics (N=39). When accounting for the fact that Lecture Notes in Computer Science compiles various venues, this result shows that AI4CBT research spreads across a wide range of journals, with no single journal standing out as specialized in AI4CBT. Therefore, it seems that AI methods are broadly integrated into CBT research venues (i.e. without being exluded to other venues) and that no AI4CBT specialist venues are standing out. This result correlates with e.g. AI for medicine research (Guo et al., 2020), in which one meta‐journal dominates the field and all other journals, which are not meta‐journals, achieve a similar level of production.

AI4CBT contents

Knowledge build‐up: citation network analysis

The contents of a scientific field can be assessed by examining how documents build on each other. Citation networks, which link documents in the corpus to those they cite or that cite them, help identify the most influential documents in a field, potential subfields, and offer insights into how knowledge is accumulated within the field.

As shown in Figure 7, several key observations emerge. The network is largely fragmented, composed of many small, disconnected subnetworks. The majority of documents (N=776) are completely isolated, with the remainder forming N=10 small clusters of two to nine documents, and one larger network comprising N=49 documents. This surprising result strongly indicates a significant lack of scientific build‐up in AI4CBT research, with 83% of AI4CBT documents neither citing nor being cited by any other AI4CBT document. The remaining documents are grouped into small, loosely connected clusters or are entirely disconnected from each other, highlighting the existence of disparate and disconnected subfields. This finding is particularly surprising given the substantial resources (e.g., expertise, networks, data, technical capability) required to engage in both AI and CBT research.

FIGURE 7.

Citation network analysis of the AI4CBT corpus. Each node represents a document in the corpus. The larger the node, the more it has been cited (in absolute terms, not only within the corpus). Each edge represents a citation relationship between two connected nodes. The figure focuses on the largest citation groups.

Key topics: keywords co‐occurrences

Keywords co‐occurrence networks are commonly used to identify the key topics within a corpus. In these networks each node represents a keyword used by the authors and each edge represents a co‐occurrence between two related keywords. The larger the node the more frequently that keyword appears and the thicker the edge the more often the two connected keywords co‐occur in the documents of the corpus. Node color indicates clusters that are generated based on sets of keywords that tend to co‐occur frequently with each other.

Figure 8 shows a keyword co‐occurrence network. To account for variations in the spelling, all forms of CBT were unified under the label “CBT", and all forms of Internet‐based CBT were unified under the label “ICBT", with plural forms converted to singular. To avoid clutter, we only retained keywords that appeared at least 10 times, reducing the vocabulary from 2261 different keywords to the 42 most commonly used ones. The plot and analysis were generated using VOSviewer (Van Eck & Waltman, 2014).

FIGURE 8.

Keyword co‐occurrence network of the keywords used by the corpus. The color of the top figure represents clusters of highly connected elements. The colors of the bottom figure represent the average year of publication of the documents featuring this keyword.

The clustering algorithm generated the following clusters, each representing a set of author keywords that frequently co‐occur:

• Cluster 1 (red): adolescent, anxiety disorders, CBT, chronic fatigue syndrome, dropout, ICBT, Internet, intervention, obsessive‐compulsive disorder, prediction, predictors, PTSD, smoking cessation, treatment, treatment outcome

• Cluster 2 (green): anxiety, chatbot, chronic pain, conversational agent, depression, digital health, mental health, mHealth, mobile health, mobile phone, obesity, stress

• Cluster 3 (blue): CBT for insomnia, deep learning, insomnia, machine learning, network meta‐analysis, panic disorder, randomized control trial, sleep, treatment response

• Cluster 4 (yellow): artificial intelligence, natural language processing, psychotherapy, virtual reality

• Cluster 5 (purple): psychosis, schizophrenia

An analysis of these clusters helps to identify overarching topics emerging from the co‐occurrence of keywords. Cluster 1 and Cluster 5 are exclusively focused on psychological and medical topics, with no AI‐related keywords. Cluster 1 also contains core CBT‐related terms, indicating converging interests around CBT. Cluster 2 reveals a convergence of interest in mobile‐based chatbots for mental health and/or negative experiences. Cluster 3 suggests a focus on machine learning applied to sleep and sleep disorders. Cluster 4 is more technologically focused, involving AI, natural language processing, and virtual reality alongside psychotherapy, indicating a strong technological focus.

The separation of topics across these clusters suggests a clear divide between core CBT‐related and AI‐related keywords. This suggests a potential division in the AI4CBT field, where research is either primarily focused on CBT or AI, but rarely both at the same time. However, some clusters integrate AI and CBT, such as chatbots and mobile technologies for anxiety, stress, and depression, as well as machine learning for insomnia. These results correlate with e.g. AI for medicine (Guo et al., 2020), which also features topics that tie strongly to the involved disciplines.

The proximity between topics can be estimated from the graphical representation in Figure 8. The visualization, which tends to position related keywords closer together (though subject to projection limitations), suggests that “CBT", “depression", and “mental health" are the most central keywords, as they are physically centered, larger, and closely connected. In contrast, technological keywords appear more scattered on the periphery and are somewhat smaller. This representation implies that AI4CBT research tends to mention applicative (CBT‐related) keywords more frequently together than AI‐related ones. This suggests that the integration of AI keywords is more scattered (i.e., documents often mention only one AI‐related keyword rather than multiple), indicating that AI in AI4CBT research may be seen as a methodological tool rather than a central focus. Combined with the cluster analysis, this finding suggests that AI4CBT is framed more as a multidisciplinary field (i.e., borrowing methods from one field to apply to another) than an interdisciplinary one (i.e., developing specialized theories, frameworks, and methods at the intersection of both fields) (Tress et al., 2005).

Lastly, the average year of occurrence for each topic provides insights into ongoing trends in the field. Topics such as artificial intelligence, mobile health, and chatbots are relatively recent and appear to be trending, while research on predictors and treatment outcomes seems to be declining. A more detailed analysis of topic trends over time would provide further insights but is left for future work due to space constraints.

AI4CBT technological landscape: terminology occurrence analysis

The keyword analysis reveals that while the vocabulary related to CBT covers a wide range of topics, the keywords tied to AI remain relatively generic. To gain a more comprehensive understanding of the technological landscape underpinning AI4CBT, Table 4 lists the number of occurrences of documents featuring various AI‐related keywords introduced in the main query, providing a broad overview of AI technologies.

TABLE 4.

Number of occurrences of documents featuring a given AI‐related keyword, as defined in the seven‐step AI‐keywords identification detailed in the method section (e.g. related literature, snowballing method). Similar terms have been merged and the number of occurrences is separated by/. Dashes and spaces can be used interchangeably without creating duplicates.

Filter name	# occurrences
Affective computing	12
Agent architecture/ agent‐based algorithm/ agent‐based comput* /agent‐	0 / 0 / 0 /4/0/
Based model* /agent‐based simulation/social simulation	0
Artificial agent/virtual agent	0/18
Artificial intelligence	162
Automated planning/markov decision process	0/1
Chatbot* /conversational agent	92/53
Chatgpt /chat gpt/prompt engineering	9/ 1/1
Decision tree	26
Expert system	2
Interactive intelligent system*	0
Large language model/bidirectional encoder representations from transformers	9/0
Logistic regression	412
Machine learning	177
Multi‐agent/multiagent	2/3
Naive bayes	1
Natural language process*	42
Neural network/backpropagation/deep learn* /generative adversarial	54/2/31 /1/
Network/perceptron/tensorflow	1/0
Random forest	28
Reinforcement learning/q‐learning	13/0
Robot*	40
Supervised learning/transfer learning	4/3
Support vector machine	18
Turing test	2

The results in Table 4 highlight several key technologies with varying levels of interest. The most frequently used method is logistic regression (N=412), followed by machine learning (N=177) and artificial intelligence (N=162). Chatbots/conversational agents and neural networks follow with about 90 occurrences each. Next are natural language processing and robotics (N∼40), followed by more specialized AI techniques such as decision trees, random forests, reinforcement learning, support vector machines, large‐language models, and ChatGPT, as well as general terms like affective computing and virtual agents (N∼15). There are near‐zero occurrences of keywords related to the agent paradigm (e.g., agent‐based modeling, social simulation, multiagent systems), automated planning, naive Bayes, Turing test, supervised learning, and transfer learning.

This layered distribution helps to structure the technological discourse and interests within AI4CBT. Aside from logistic regression, a specific algorithm already well integrated into CBT‐related fields (e.g., psychology and medical sciences), most technological discussions are dominated by broad terms such as AI, machine learning, and neural networks, rather than specialized techniques like Markov decision processes or support vector machines. This suggests that AI may be used in a more general or evocative manner (e.g., “machine learning was used to process the results") rather than employing specific methods. Furthermore, the use of non‐specialized terminology implies that the intended audience may be less focused on the technological details. This observation hints at the absence of streamlined AI methods specialized for AI4CBT research, which may represent an important next step for the field. Additionally, the moderate representation of chatbots, conversational agents, and robots in the corpus might indicate a moderate interest in AI‐powered interfaces.

These results corroborate findings from (Aggarwal et al., 2023; Guo et al., 2020; Li et al., 2023), which highlighted the use of terms like machine learning and artificial intelligence in AI4CBT research and in AI for medicine more generally. However, except for AI for medicine, these former studies differ regarding key terms, such as logistic regression, which was less emphasized in previous literature. It can be hypothesized that logistic regression may be less used in clinical settings, as suggested by (Jiang et al., 2024).

From a historical perspective, this analysis shows that AI4CBT heavily focuses on techniques that (re)emerged and gained significant attention after 2015 (e.g., machine learning, neural networks, natural language processing), while being relatively disconnected from classical AI methods (e.g., agent paradigms, automated planning, Bayesian models, and expert systems). This is noteworthy, as many of these traditional methods could directly address challenges in CBT delivery (e.g., automatic planning for organizing CBT modules or allocating patients to therapists), offering exciting prospects for future research if these AI methods become more widely adopted by CBT practitioners.

Disciplinary transfers: citation analysis

This section provides a citation analysis by studying the citation relationships between disciplines. Citations‐in refer to the references made by documents in the corpus (i.e., from the reference lists of the corpus documents), and citations‐out refer to documents that cite the corpus (i.e., the list of documents retrieved using the “cited‐by" function in Google Scholar).

Figure 9 presents the proportion of citations‐in and citations‐out by discipline, showing which disciplines are citing and being cited by other disciplines. ⁶

FIGURE 9.

Sankey diagram representing the flow of citations between the disciplines of the corpus. The central column represents the corpus, the left column represents citations‐in (documents cited by the corpus), and the right column represents citations‐out (documents that cite the corpus). Each rectangle corresponds to a discipline. For example, the purple rectangle on the left represents all citations‐in from mind sciences (i.e., psychology, neurology), the pink rectangle in the middle represents the number of citations‐in from medical sciences, and the blue rectangle on the right represents the number of citations‐out from computational sciences documents. The edges indicate the size of the citation flows. For instance, the pink edge connecting the left purple rectangle to the central pink rectangle represents the proportion of citations‐in from medical sciences to mind sciences. Based on the edge size, medical sciences account for about 30% of all citations‐in for mind sciences documents in the corpus.

The first key observation is that the three core disciplines in AI4CBT (i.e., medical sciences, mind sciences, and computational sciences) consistently cite and are cited by medical and mind sciences. This is unsurprising, given that AI4CBT themes naturally involve these two disciplines. However, computational sciences are negligibly cited by medical and mind sciences. This is particularly surprising, as one would expect computational sciences, where most AI methods are developed, to be highly cited in AI4CBT research.

CONCLUSION

This paper provides a comprehensive overview and mapping of the scientific status of the AI4CBT field by analyzing the metadata of over 50000 documents, 4616 authors, and 1671 institutions. An array of bibliometric measures was applied to assess the field, including general bibliometric figures, trends over time, publication venues, publishing disciplines, contributors at the national, institutional, and individual levels, co‐authorship networks, keywords, AI methods, and disciplinary citation analyses. This broad‐spectrum analysis offers an overview of the main trends regarding productivity, producers, publication avenues, and research content.

The findings build upon and extend previous insights from systematic literature reviews and bibliometric analyses (Aggarwal et al., 2023; Boucher et al., 2021; Higgins et al., 2023; Jiang et al., 2024). Overall, many of the results align with those of (Chen et al., 2024), including the upward trend, key disorders, core technologies, and community structure. A notable observation is the divergence in topic framing compared to top‐down perspectives in systematic reviews, which often focus on either CBT activities or performance criteria for AI systems. This highlights the relevance of both approaches and suggests a potential discrepancy between how AI4CBT documents are presented and the structures researchers aim to establish.

As identified in the results, topics of interest in AI4CBT papers primarily revolve around disorder types (e.g., anxiety disorders, depression), CBT delivery methods (e.g., ICBT), and technologies (e.g., chatbots). The identified focus on depression and anxiety disorders, which further aligns with findings from (Aggarwal et al., 2023; Chen et al., 2024; Li et al., 2023), shows a specific interest in the AI4CBT community, given that classic CBT research identifies more disorders as core interests (Kariri & Almubaddel, 2024). This specific interest correlates with the introduction of AI is not exclusive to AI4CBT as similar findings can be found in AI and medicine (Guo et al., 2020). Such an interest can be explained as arising from community factors (AI methods spread within a specialist community organized around a given ailment) and technical factors (AI methods achieving good performance addressing ailment‐specific needs). Our results diverge from prior literature by identifying regression analysis as a key method, likely due to varying definitions of AI methods in previous studies. Delivery methods, particularly regarding ICBT, were not highlighted in related work, potentially due to cut‐off dates in 2023, further confirming the emerging trend of AI in ICBT research.

Summary

In summary, AI4CBT is a rapidly growing field, and interest is growing relatively faster than both its mother disciplines (i.e. AI and CBT). AI4CBT is produced predominantly by research‐leading countries, with notable prevalence in the US and a convergence of strong institutions in Sweden. At the level of individual contributors, the impressive number of contributors is seemingly vastly composed of a majority of low producers with high turnover (a fifth of a document in average) while a relatively large community of top contributors show similar levels of contributions, suggesting a healthy level of redundancy and hence robustness of the field. Despite this redundancy, the AI4CBT community remains loosely connected, with researchers and research tracks often operating in near isolation from each other despite relatively large author teams. Likewise, AI4CBT research is distributed across a variety of journals, without a specialized publication venue.

A focus on contents indicates that whereas AI4CBT seems primarily published along the three core disciplines one may expect from the topic (i.e. computational science, mind sciences, and medical sciences), the interdisciplinary integration surprisingly seems to remain shallow. Computational sciences are seemingly only marginally referenced within non‐computational science AI4CBT contributions. Moreover, the introduction of AI in documents seems to be primarily performed through imprecise terminologies despite AI offering of specific methods, except for logistic regression –which a method already well‐integrated by non‐computational science disciplines and has the benefit of being more straightforward to interpret. Therefore, it seems that AI4CBT is, currently, primarily multidisciplinary (i.e., applying standard AI methods in a CBT context) rather than interdisciplinary (i.e., developing AI methods specialized for CBT concerns) (Tress et al., 2005).

Overall, the findings suggest that AI4CBT fits the profile of an emerging scientific field by meeting the five key characteristics identified by (Rotolo et al., 2015): radical novelty (i.e., the technology achieves a function using a different principle than previously used); fast growth (particularly since 2017, with contributions from about 5000 authors and nearly 1000 documents); coherence‐creating activity (i.e., uniting previously separate research fields); prominent impact (e.g., approximately 5% of all CBT research now involves AI4CBT, with continued growth); and uncertainty and ambiguity (i.e., the application of AI4CBT is still under consolidation). The field has not yet reached the post‐emergence phase characterized by an S‐shaped production curve over time (Mansfield, 1961; Price, 1963; Rogers et al., 2014). AI4CBT can also be described as in a pre‐paradigmatic stage, lacking a cohesive framework and characterized by competing schools of thought (Shapere, 1964), as evidenced by the loose topology of its co‐authorship and co‐citation networks (Bettencourt et al., 2009). A strong parallel can be made with the trends observed in the emerging field of AI for medicine research (Guo et al., 2020) (e.g., rapid uptake, mainly by high‐income countries, focus on a subset of disorders, reliance on a few AI methods that dominate the field).

Challenges and opportunities of the field

AI4CBT can be expected to be subject to the structural challenges and opportunities commonly attributed to the emerging, pre‐paradigmatic, interdisciplinary, decentralized phase it seems to belong to, as identified in the previous section. Such a decentralized development inherently breeds a low level of organization in the field, which can be expressed as various research interests, communities, and publication venues evolving in parallel. While natural in the early days of a field is in its cradle, such disorganization creates increasing challenges as the field grows. First, disorganization creates inefficiencies: AI4CBT requires significant resources, including expertise, collaboration with scientists and practitioners, data, technical skills, and software development. This demand for resources is multiplied by the number of independent groups developing parallel tracks, resulting in reduced productivity as efforts are duplicated. Second, as AI4CBT grows, so does its societal impact, raising concerns that need to be addressed by the community. What is the impact of AI in CBT on social, legal, ethical, organizational, medical, cultural, political, environmental, economic, and technological domains? How can such socio‐technical systems be responsibly designed? While related work has touched on these issues (Aggarwal et al., 2023; Boucher et al., 2021; Higgins et al., 2023; Jiang et al., 2024; Li et al., 2023), there seems to be growing concern that these topics are not being adequately addressed, a problem seen in both mental health digitalization (Sucala et al., 2017; Wasil et al., 2021) and AI (Lindgren, 2023a; Lindgren, 2023c). The history of technological lock‐ins shows how pervasive and harmful these issues can be. To ensure the continued success of the field and to serve the broader goals of mental health and society, the community must begin addressing these challenges.

A straightforward path forward involves establishing AI4CBT as an organized entity, with development along two main lines: social and scientific. Socially, the field can grow in multiple areas (Shneiderman, 2016; Vanhée, 2023), including establishing a clear identity, defining visions and boundaries, building affiliated communities (e.g., special interest groups, workshops, conferences), creating venues for merging research threads (e.g., special issues, specialized books), developing methodological tools (e.g., specialized methods, reusable software, quality standards), creating operational resources (e.g., reusable platforms, datasets, networks), designing training programs, creating sustainable career paths, and streamlining outreach to practitioners. To achieve its full potential for social impact, it seems essential for AI4CBT research and methods to reach out low‐income areas through, e.g. international funding, research partnerships, and openly access AI4CBT tools. While these areas broadly miss out on the AI4CBT research landscape, they also feature the greatest exposure to mental health harm and the least accessibility to low‐cost mental healthcare. As such, AI4CBT, by allowing to further streamline the prospects offered by CBT, has a very strong leverage for achieving social impact (Santiago et al., 2013). Likewise, the social impact of AI4CBT applications may be raised by exploring a wider diversity of disorders, beyond anxiety disorders which overrepresentation stands out in the results.

Scientifically, the field must establish a unified vocabulary, frameworks, and theories (Jerneck et al., 2011), while continuing and expanding regular review work (Aggarwal et al., 2023; Boucher et al., 2021; Higgins et al., 2023; Jiang et al., 2024; Li et al., 2023). To realize its full potential, AI4CBT must evolve beyond its current multidisciplinary approach, which primarily involves the seeming shallow use of off‐the‐shelf AI methods to CBT research. Instead, the field would strongly benefit from embracing interdisciplinarity (i.e., developing specialized ontologies, theories, and applications at the intersection of the involved disciplines) or even transdisciplinarity (i.e., interdisciplinary research that involves non‐academic stakeholders) (Tress et al., 2005). AI4CBT has the potential to foster the design of innovative AI systems that prioritize human factors, benefiting both AI and CBT research. This could involve leveraging currently underutilized technologies related to CBT factors, such as affective computing and cognitive modeling (Espie, 2007; Vanhée et al., 2021; Zucco et al., 2017), automated planning (Baars & Van Merode, 2008; Lindsay et al., 2024; Vanhée et al., 2022), agent‐based models in social simulation (Horned & Vanhée, 2023a; Kreulen et al., 2022; Squazzoni et al., 2020), psychosocial models (Castro Martínez & Santamaría‐García, 2023; Garcia et al., 2023; Horned & Vanhée, 2023b), game integration (Arenas et al., 2022; Szczepanska et al., 2022; Wilkinson et al., 2008), as well as addressing AI‐related risks such as bias, dehumanization, and privacy concerns (Salim & Yusoff, 2008; Skorburg et al., 2024), while training the next generation of AI designers (Borenstein & Howard, 2021; Garrett et al., 2020; Vanhée & Borit, 2022).

Limitations and future work

Bibliometric analyses, like any other method, inherently come with limitations in the results they can produce (Haustein & Larivière, 2014). Bibliometric databases are incomplete and may miss certain documents, potentially introducing systematic biases (e.g., fields that publish in venues not indexed by Scopus). Keyword‐based querying may cause relevant documents to be missed when these documents do not explicitly mention the exact keywords. These missing documents can be expected to be unproblematic in the current case, as there is no reason to believe that the complete set would be statistically different from the current one except for larger numbers in the overview. Additional terms can be considered (e.g., specific therapeutic methods), but, unless having an exhaustive listing of these terms and none of them involve false positives, their introduction may introduce biases towards specific concepts and conceptual frames. Similarly, any categorical representation of disciplines is inherently reductionist, and Scopus' discipline attribution, based on expert input, may be subject to inconsistencies. Furthermore, bibliometric tools offer a limited perspective, constrained by the metadata they can retrieve, which may overlook important elements like hidden citations (Meng et al., 2024). Given that bibliometric analyses rely heavily on statistics, they may sometimes produce misleading results due to statistical anomalies –particularly when conducting the kind of broad overview presented in this paper. While we believe the results reflect the collected evidence, the combination of these factors calls for cautious interpretation before drawing definitive conclusions. Several measures were taken to mitigate these risks, including an advanced data quality analysis to rule out false positives and false negatives, complementary analyses to triangulate surprising findings, and extensive manual checks of the corpus content and relevance of the results. The uncertainty surrounding the exact distinction between medical and mind sciences in Scopus is non‐critical, as both disciplines demonstrated similar trends. Thus, we are relatively confident that the broad overview presented in this paper aligns with the practical reality of the AI4CBT field.

For future work, the broad overview conducted in this paper can be further refined with additional analyses focusing on specific variables of interest. A comprehensive content analysis could be performed to identify the key topics of the corpus beyond the authors' keywords. Such a content analysis could identify trends of topics over time and leverage advanced topic modeling techniques, such as Latent Dirichlet Allocation (Blei et al., 2003), which has proven effective in related fields for identifying key topics (Syed et al., 2018). Additionally, a further analysis could examine the trends over time of various topics and methods to determine whether AI4CBT follows broader AI research trends, what technological paths remain active, and why others may have fallen out of favor. This could also provide general insights for the field (e.g., revisiting earlier findings in light of new methods, reanalyzing databases). Moreover, as we identified the presence of academic silos and parallel tracks within the literature, a similar method could be applied to investigate how closely aligned or distinct the issues addressed by these silos are, as seen in (Javed et al., 2022). Such an investigation could reveal opportunities for collaboration across silos that may be duplicating efforts, enabling more efficient progress by reducing redundant work. A specialized disciplinary analysis could also be undertaken to examine the interdisciplinary profiles of author teams and research silos. This would help determine how disciplinary boundaries may be contributing to the fragmentation of the field and provide a basis for connecting experts in application, technology, and social impact to develop AI4CBT technologies that integrate psychological and clinical grounding, technical precision, and social relevance.

EXHAUSTIVE QUERY

(((TITLE‐ABS["affective computing"]) OR (SRCTITLE["affective computing"]) OR (AUTHKEY["affective computing"]) OR (TITLE‐ABS["agent architecture"]) OR (SRCTITLE["agent architecture"]) OR (AUTHKEY["agent architecture"]) OR (TITLE‐ABS["agent‐based algorithm"])

OR (SRCTITLE["agent‐based algorithm"]) OR (AUTHKEY["agent‐based algorithm"]) OR

(TITLE‐ABS["artificial agent"]) OR (SRCTITLE["artificial agent"]) OR (AUTHKEY["artificial agent"]) OR (TITLE‐ABS("artificial intelligence")) OR (SRCTITLE("artificial intelligence")) OR (AUTHKEY("artificial intelligence")) OR (TITLE‐ABS["automated planning"]) OR (SRCTITLE["automated planning"]) OR (AUTHKEY["automated planning"]) OR

(TITLE‐ABS["backpropagation"]) OR (SRCTITLE["backpropagation"]) OR (AUTHKEY["backpropagation"]) OR (TITLE‐ABS["bayesian network"]) OR (SRCTITLE["bayesian network"]) OR (AUTHKEY["bayesian network"]) OR (TITLE‐ABS["bidirectional encoder representations from transformers"])

OR (SRCTITLE["bidirectional encoder representations from transformers"]) OR (AUTHKEY["bidirectional encoder representations from transformers"]) OR (TITLE‐ABS["chat gpt"]) OR (SRCTITLE["chat gpt"]) OR (AUTHKEY["chat gpt"]) OR (TITLE‐ABS["chatbot*"])

OR (SRCTITLE["chatbot*"]) OR (AUTHKEY["chatbot*"]) OR (TITLE‐ABS["chatgpt"]) OR (SRCTITLE["chatgpt"]) OR (AUTHKEY["chatgpt"]) OR (TITLE‐ABS["conversational agent"])

OR (SRCTITLE["conversational agent"]) OR (AUTHKEY["conversational agent"]) OR

(TITLE‐ABS["decision tree"]) OR (SRCTITLE["decision tree"]) OR (AUTHKEY["decision tree"])

OR (TITLE‐ABS["deep learn*"]) OR (SRCTITLE["deep learn*"]) OR (AUTHKEY["deep learn*"]) OR (TITLE‐ABS["expert system"]) OR (SRCTITLE["expert system"]) OR (AUTHKEY["expert system"]) OR (TITLE‐ABS["generative adversarial network"]) OR (SRCTITLE["generative adversarial network"]) OR (AUTHKEY["generative adversarial network"]) OR (TITLE‐ABS["interactive intelligent system*"]) OR (SRCTITLE["interactive intelligent system*"]) OR (AUTHKEY["interactive intelligent system*"]) OR (TITLE‐ABS["large language model"]) OR (SRCTITLE("large

language model")) OR (AUTHKEY["large language model"]) OR (TITLE‐ABS["logistic regression"]) OR (SRCTITLE["logistic regression"]) OR (AUTHKEY["logistic regression"]) OR (TITLE‐ABS("machine learning")) OR (SRCTITLE("machine learning")) OR (AUTHKEY("machine learning")) OR (TITLE‐ABS["markov decision process*"]) OR (SRCTITLE["markov decision process*"]) OR (AUTHKEY["markov decision process*"]) OR (TITLE‐ABS["naive bayes"]) OR (SRCTITLE["naive bayes"]) OR (AUTHKEY["naive bayes"]) OR (TITLE‐ABS["natural language process*"]) OR (SRCTITLE["natural language process*"]) OR (AUTHKEY["natural language process*"]) OR (TITLE‐ABS["neural network"]) OR (SRCTITLE["neural network"]) OR (AUTHKEY["neural network"]) OR (TITLE‐ABS["perceptron"]) OR (SRCTITLE["perceptron"])

OR (AUTHKEY["perceptron"]) OR (TITLE‐ABS["prompt engineering"]) OR (SRCTITLE["prompt engineering"]) OR (AUTHKEY["prompt engineering"]) OR (TITLE‐ABS["q‐learning"]) OR (SRCTITLE["q‐learning"]) OR (AUTHKEY["q‐learning"]) OR (TITLE‐ABS["random forest"]) OR (SRCTITLE["random forest"]) OR (AUTHKEY["random forest"]) OR (TITLE‐ABS["reinforcement learning"]) OR (SRCTITLE["reinforcement learning"]) OR (AUTHKEY["reinforcement learning"]) OR

(TITLE‐ABS["supervised learning"]) OR (SRCTITLE["supervised learning"]) OR (AUTHKEY["supervised learning"]) OR (TITLE‐ABS["support vector machine"]) OR (SRCTITLE["support vector machine"]) OR (AUTHKEY["support vector machine"]) OR (TITLE‐ABS["tensorflow"]) OR (SRCTITLE["tensorflow"]) OR

(AUTHKEY["tensorflow"]) OR (TITLE‐ABS["transfer learning"]) OR (SRCTITLE["transfer learning"]) OR (AUTHKEY["transfer learning"]) OR (TITLE‐ABS["turing test"]) OR (SRCTITLE["turing test"]) OR (AUTHKEY["turing test"]) OR (TITLE‐ABS["virtual agent"]) OR (SRCTITLE["virtual agent"]) OR (AUTHKEY["virtual agent"])) OR ((TITLE‐ABS["agent‐based comput*"]) OR (SRCTITLE["agent‐based comput*"]) OR (AUTHKEY["agent‐based comput*"]) OR (TITLE‐ABS["agent‐based model*"]) OR (SRCTITLE["agent‐based model*"]) OR (AUTHKEY["agent‐based model*"]) OR (TITLE‐ABS("agent‐based

simulation")) OR (SRCTITLE["agent‐based simulation"]) OR (AUTHKEY["agent‐based simulation"]) OR (TITLE‐ABS["agentbased comput*"]) OR (SRCTITLE["agentbased comput*"]) OR (AUTHKEY["agentbased comput*"]) OR (TITLE‐ABS["agentbased model*"]) OR (SRCTITLE["agentbased model*"]) OR (AUTHKEY["agentbased model*"]) OR (TITLE‐ABS["agentbased simulation"]) OR (SRCTITLE["agentbased simulation"]) OR (AUTHKEY["agentbased simulation"]) OR (TITLE‐ABS("individual‐based

model*")) OR (SRCTITLE["individual‐based model*"]) OR (AUTHKEY["individual‐based model*"]) OR (TITLE‐ABS["multi‐agent"]) OR (SRCTITLE["multi‐agent"]) OR (AUTHKEY["multi‐agent"])

OR (TITLE‐ABS["multiagent"]) OR (SRCTITLE["multiagent"]) OR (AUTHKEY["multiagent"]) OR (TITLE‐ABS["social simulation"]) OR (SRCTITLE["social simulation"]) OR (AUTHKEY["social simulation"])) OR ((TITLE‐ABS["robot*"]) OR (SRCTITLE["robot*"]) OR (AUTHKEY["robot*"]))) AND ((TITLE‐ABS["cognit* behav* interv*"]) OR (SRCTITLE["cognit* behav* interv*"]) OR (AUTHKEY["cognit* behav* interv*"]) OR (TITLE‐ABS["cognit* behav* psychotherap*"]) OR (SRCTITLE["cognit* behav* psychotherap*"]) OR (AUTHKEY["cognit* behav* psychotherap*"]) OR (TITLE‐ABS["cognit* behav* therap*"]) OR (SRCTITLE["cognit* behav* therap*"])

OR (AUTHKEY["cognit* behav* therap*"]) OR (TITLE‐ABS["cognit* behav* treat*"]) OR (SRCTITLE["cognit* behav* treat*"]) OR (AUTHKEY["cognit* behav* treat*"]))

WORLD MAP

A worldmap provides an illustrative oversight of the the degree of development of AI4CBT in various geographic areas.

Vanhée, L. , Andersson, G. , Garcia, D. , & Sikström, S. (2025). The rise of artificial intelligence for cognitive behavioral therapy: A bibliometric overview. Applied Psychology: Health and Well‐Being, 17(2), e70033. 10.1111/aphw.70033

DATA AVAILABILITY STATEMENT

Data sharing not applicable.