A review of artificial intelligence enhanced cognitive behavioural therapy using the BECK AI BOT for mental health interventions

Abstract

The integration of artificial intelligence (AI) and cognitive behaviour therapy (CBT) is a revolutionary solution to the global mental health issue, characterized by increasing need and decreased access to treatment. This research investigates the potential of AI-fortified cognitive behavioural therapy technologies, including chatbots, virtual reality, and adaptive learning modules, to enhance the efficacy, accessibility, and individualization of treatment for anxiety, depression, and PTSD. The study evaluates the scalability, ethical issues, and therapeutic efficacy of the therapies by combining peer-reviewed and experimental data. The suggested methodology combines AI-driven conversational therapy with predictive modelling to deliver individualized, real-time mental health treatment. In this study, a conceptual chatbot prototype, designated BECK-AI BOT, was developed to illustrate the application’s interface and functionality, enhancing accessibility for both patients and therapists in the future. This study does not present new clinical trial data. All reported symptom-reduction and engagement findings are drawn from previously published studies of existing AI-driven CBT systems (e.g., Woebot, Wysa, Eleos, Limbic). The present work offers a narrative synthesis of current evidence and introduces a conceptual architecture and prototype (BECK-AI BOT), without evaluating it clinically. Notwithstanding these difficulties, problems persist, including a lack of long-term efficacy statistics, cultural sensitivity issues, and moral reservations about over-reliance on AI during emergencies. The argument comes in the form of AI possibly improving, not replacing, human therapists, emphasizing hybrid systems for fair treatment. Future research needs to advance emotional intelligence within AI, which combines AI-driven conversational therapy and predictive modelling to deliver real-time, personalized mental health services.

Introduction

Cognitive Behaviour Therapy (CBT) has long been a cornerstone of evidence-based psychotherapy and has effectively treated common mental disorders such as anxiety, depression, and post-traumatic stress disorder (PTSD) [1]. CBT previously relied on scheduled meetings with experienced therapists and consisted of cognitive restructuring and behavioural interventions. However, the recent rapid rise in global mental health challenges, because of post-pandemic mental trauma, digital overwhelm, and socio-economic uncertainty, has created a pressing need for more rapid, more accessible, and scalable mental health interventions [2]. The shortage of skilled therapists, particularly in low-resource settings, also increases the treatment gap. In this regard, applying Artificial Intelligence (AI) to CBT has the potential to render mental care democratically accessible. AI-based instruments, like online platforms and robots, can treat clients who may otherwise be bereft of such treatment due to geography, economics, or culture [3]. This intersection of AI and psychology could potentially revolutionize mental healthcare by providing evidence-based interventions at scale that are responsive and tailored to user requirements in real-time. Recent work on computational psychotherapy systems further demonstrates that conversational agents equipped with theory-of-mind (ToM) simulation modules can more accurately model users’ emotional states and predict behavioural change, achieving higher adaptiveness than earlier chatbots, such as Woebot or Wysa [4].

The efficacy results referenced in this paper describe findings from prior peer-reviewed studies of AI-supported CBT platforms. As no empirical or clinical data were generated for the BECK-AI BOT itself, this manuscript should be understood as a literature-based narrative review, supplemented by a proposed technical framework. It applies a sample of recent peer-reviewed literature, including meta-analyses and systematic reviews, from electronic health databases and peer-reviewed journals. The review focuses on AI-fortified CBT interventions involving conversational agents (e.g., Woebot, Wysa, and Youper), VR exposure therapy, and adaptive learning modules simulating therapeutic activities [3]. The assessment intervention continuum described in digital mental-health research shows that smartphone-based monitoring, ecological momentary assessment (EMA) and just-in-time adaptive interventions can substantially enhance timely detection and support for stress, anxiety and depression, underscoring the need for AI systems that integrate continuous assessment with personalised therapeutic responses [5]. The studies were selected based on methodological quality, sample heterogeneity, and relevance to personalization, scalability, clinical efficacy, and ethics. Such a strategy enables broad sensitivity to the extent to which AI-enhanced CBT products are being used, tested, and adopted in clinical and non-clinical settings. Literature was thoroughly reviewed for outcomes that range from symptom reduction, user engagement, therapeutic alliance, and usability. Besides, the emphasis was placed on whether such interventions aligned with normative CBT principles or deviated from therapeutic standards.

Gaps in the research

Though research on CBT enabled through AI is increasing at a high rate, the research gaps remain huge, and they should be addressed so that deployment can be successful and ethical. One of the major weaknesses is the absence of an intense investigation of long-term therapeutic effects and the continuity of therapeutic relationships formed with AI systems. Although short-term relief from symptoms is promising, it is unclear how long the effects last. Also, most of the present models were trained and tested upon Western populations, and consequently, the applicability of such models to multicultural and socioeconomic settings is minimal [6] There aren’t robust mechanisms for addressing high-risk or crises, an intrinsic part of mental health treatment that existing AI technology isn’t well-suited to handle. Perspectives from persuasive-technology research further emphasise that digital mental-health systems can help reduce global inequities by addressing treatment gaps, economic barriers and limited service availability, particularly in alignment with SDG-driven priorities for accessible behaviour-change interventions [7]. Ethical considerations of data protection, user control, and algorithmic fairness further complicate the application of such tools to vulnerable groups [8]. Such lacunae necessitate more comprehensive, longitudinal, and ethically robust studies to assess both strengths and constraints of AI-based CBT treatment.

Scope of the survey

The scope of this review encompasses the broad range of AI-driven CBT software under consideration with a focus on their therapeutic efficacy, usability, and use in clinical settings. The review presents an overview of chatbot-guided treatments like Woebot, Wysa, and Youper that have already demonstrated remarkable effects on depression and anxiety symptom reduction across diverse samples. Notably, Woebot has been highly interactive and active among its users, making it an alternative suitable for individuals who prefer not to use traditional therapy [8]. Review also thinks about virtual reality-augmented CBT interventions with the point of exposure treatment in situations where there is PTSD, OCD, as well as anxiety disorders. Virtual reality-based apps recreate real-life environments where patients can expose their aversions incrementally and safely, leading to measurable symptom reduction [8,9]. Lastly, AI systems that incorporate adaptive learning algorithms to facilitate self-guided CBT modules are assessed for their ability to personalize therapeutic content. Recent work in intelligent cognitive systems also highlights the value of panel datasets that combine diagnostic questionnaires with daily diary inputs to support personalised forecasting and behavioural modelling in stress, anxiety and depression [10]. A comprehensive review aims to give a multi-aspect view of how AI tools are being applied to augment or supplement partially human-delivered CBT interventions.

Objectives of the paper

The main objective of this paper is to critically examine how artificial intelligence can be integrated into Cognitive Behavioural Therapy (CBT) to improve personalisation, accessibility and ethical practice. Prior studies demonstrate that AI systems can support early detection of anxiety, depression and suicidal ideation through linguistic modelling and annotated cognitive frameworks [11], while also delivering personalised CBT techniques that lead to short-term symptom improvement [12,13]. A further aim is to assess user satisfaction and the therapeutic relationship in AI-supported interventions, given evidence that engagement and perceived alliance play a central role in digital-CBT adherence [14,12]. The paper also evaluates whether AI tools function best as complementary supports to clinicians or as scalable solutions in low-resource settings, as indicated in hybrid human–AI models and AI-assisted supervision studies [15,16]. Finally, the study seeks to provide a balanced overview of the advantages and limitations of AI-based mental-health interventions, considering both their potential for increased reach and their ethical challenges, including safety, bias and governance [17,18].

• i.Evaluate the Effectiveness of AI-enhanced CBT Tools.
• ii.Enhance Accessibility and Affordability of Mental Health Services.
• iii.Personalize Psychological Interventions like CBT Using AI Algorithms.
• iv.Explore the Role of AI in Therapist Training and Decision Support.
• v.Investigate Ethical, Emotional, and Societal Impacts of AI in Psychotherapy.

Structure of the paper

The paper explores the integration of Artificial Intelligence (AI) into Cognitive Behavioral Therapy (CBT) to enhance mental health care by improving accessibility, personalization, and efficiency. Section  1 begins by introducing the growing mental health crisis and the potential of AI tools like chatbots and virtual reality to bridge treatment gaps. The study outlines key objectives, including evaluating the clinical effectiveness of AI-driven CBT, its role in therapist support, and ethical concerns. A detailed literature review and research table summarize past studies, followed by experimental reviews highlighting real-world applications and outcomes.

The remainder of this paper is structured into six sections. Section 2 introduces the technology behind AI in CBT, i.e., natural language processing (NLP), deep learning, and adaptive algorithmic models that simulate therapeutic conversation and feedback. Section  3 covers varied AI-based tools like chatbots and virtual reality platforms, detailing their construction, functionality, and integration within CBT models. Section  4 describes the proposed model of BECK-AI BOT. mentions existing limitations in research, like restricted cultural diversity in training data and longitudinal testing requirements. Section  5 addresses how clinical safety and ethics is followed and regulated, since the prototype is not deployed in real world setting a scenario of guidelines and principles that should be followed in the post production framework is given. Section  6 mentions existing limitations in research, like restricted cultural diversity in training data and longitudinal testing requirements. It also talks about personalization, accessibility, ethics, and clinical integration. Finally, Sect.  7 outlines the findings and recommends future research and application of clinically effective and socially equitable AI-facilitated CBT interventions.

Literature review

A structured narrative search was conducted in PubMed, PsycINFO, Scopus, and Google Scholar for papers published between 2015 and 2024, using keywords such as “AI in mental health,” “AI-enhanced CBT,” “digital mental-health intervention,” and “mental-health chatbot.” Editorials, opinion articles, non-English publications, and duplicates were rejected. Studies that looked at AI-assisted CBT, reported clinical, engagement, or system-performance outcomes, or detailed pertinent technological frameworks were included. Randomised trials, pilot studies, observational research, and technical evaluations involving adults or adolescents with anxiety, depression, stress, or PTSD were all acceptable study designs. To ensure consistent synthesis and inform the proposed BECK-AI BOT framework, outcomes were categorized as clinical, focusing on symptom improvement, user patterns, and machine learning metrics that address the value of accuracy.

Integrating artificial intelligence (AI) into cognitive behavioural therapy (CBT) has progressed across several modalities, each contributing important insights for personalisation, accessibility and system design. The first major area of development involves AI-driven conversational and socially assistive agents. Dino et al. explored the use of a socially assistive robot, Ryan, for delivering internet-based CBT (iCBT) to depressed older adults. Using affect-sensitive interactions and AIML-based dialogue, the study demonstrated improved engagement, emotional response and symptom reduction over four weeks, showing that robot-delivered CBT may be particularly beneficial for socially isolated populations [19] as shown in Table 1. This supports the need for emotionally attuned conversational strategies in systems such as the proposed BECK-AI BOT. Similarly, McFadyen et al. (2024) assessed a generative AI-powered CBT app Limbic Care in Table 2, which offered personalised interventions via an AI chatbot. Users interacted 2.4 times more frequently and for 3.8 times longer than those using static materials, alongside greater symptom relief and no increased risk of adverse events, demonstrating scalability and safety in routine mental-health use [18]. These findings highlight the importance of integrating adaptive engagement and monitoring functions, elements reflected in the BECK-AI BOT’s EMA and TCN forecasting components.

Table 1.

Observational and engagement-focused studies of AI-supported CBT and chatbots

Reference	N (Sample Size)	Intervention	Study Design	Population	Duration	Outcome	Behavioural Change	Advancements in the current study
Manole et al. [12]	50	ChatGPT-based AI chatbot delivering CBT-informed anxiety management	Two-phase observational study	Adults with mild–moderate anxiety (BAI or GAD-7 screened)	Phase 1: 7 days; Phase 2: 7 days, two months later	Clinical and engagement	Significant reductions in anxiety Symptoms: mean improvement ≈ 21.15% (Phase 1) and 20.42% (Phase 2); participants reported high satisfaction and sustained engagement, but reported limited empathy compared with humans.	Our review goes beyond symptom tracking by integrating EMA-based adaptive personalisation, MentalBERT-based cognitive analysis, TCN forecasting, crisis-safety layers and clinician oversight, which Manole’s study does not include.
Omarov et al. [17]	N.A	AIML-based mobile CBT chatbot	User studies: pre-post evaluation	Adults with varying levels of psychological distress	Short-term use with several sessions/interactions	Clinical and engagement	Participants reported significant improvements in perceived mental well-being and positive feedback on accessibility and personalisation, The The evidence is preliminary and based on self-report.	Advances beyond preliminary self-report evidence by providing a reproducible architecture (RASA + BERT + TCN), explicit safety governance, and integration of predictive modelling absent in the AIML prototype.
Frischholz et al. [14]	Small adult sample	Virtual agent delivering CBT-based exercises	Experimental pre-post study	Adults with elevated stress/anxiety symptoms	Several sessions across a brief programme	Clinical and engagement	Reported reductions in distress and improved mood and Participants rated the virtual agent as helpful and acceptable.	Goes beyond single-function virtual agents by incorporating deep-learning NLP (MentalBERT), safety protocols, and predictive symptom modelling, creating a more robust clinical-use framework.
Dino et al. [19]	N.A	Conversational social agent implementing CBT for low mood	Pilot pre-post study	Adults with depressive symptoms	Multi-week CBT-style conversational programme	Clinical and Engagement	Pre-post improvements on depression and anxiety scales with high levels of engagement and with the conversational agent	Builds on preliminary pilot findings by proposing a fully integrated architecture combining CBT delivery, NLP, predictive analytics, EMA tracking, and human clinician monitoring.
Creed et al. [16]	Samples differ in each phase	Lyssn platform providing AI-generated CBT fidelity metrics and supervision tools	Implementation or service study	Community mental health therapists and their CBT caseloads.	Phase based	Engagement or Service outcome	Demonstrated feasibility and usability of AI-supported supervision, AI fidelity metrics approached ≥ 80% of human reliability and potential to sustain CBT quality and support large-scale implementation.	Expands beyond clinician support tools by integrating both clinician-facing and user-facing components (chatbot, forecasting, safety), forming a holistic ecosystem rather than supervision-only augmentation.
Hassan et al. [20]	Large Clinical sample	AI analytics is integrated into routine CBT, which includes symptom tracking and outcome dashboards.	Naturalistic observational study	Real-world CBT clients in routine care	Routine care over weeks to months	Clinical and Service Outcomes	Reported improved routine outcome monitoring and possibly better symptom trajectories when AI-enabled measurement was used.	Advances beyond analytics dashboards by providing a comprehensive CBT intervention system with real-time NLP, crisis detection, personalised pacing, and TCN-based relapse prediction.

Table 2.

Clinical trials and controlled studies of AI-supported CBT and chatbots

Author	N	Intervention	Design	Population	Control Group	Duration	Outcome	Key Results
Fulmer et al. [3]	75	Tess psychological AI chatbot delivering CBT-informed support	Randomised controlled trial	US university students with self-identified depression/anxiety	Two Tess access conditions (2 vs. 4 weeks) vs. information-only e-book	2–4 weeks post-intervention	Clinical	Tess’s daily check-in group showed significantly greater reduction in depressive Symptoms (PHQ-9) vs. control and significant reduction in anxiety (GAD-7) In both Tess groups.
Sadeh-Sharvit et al. [15]	47	AI platform augmenting CBT sessions (automatic transcription, feedback on CBT techniques, progress notes)	Randomised clinical trial (AI-supported CBT vs. treatment-as-usual)	Adult outpatients with depressive or anxiety disorders in a US community clinic	CBT with Eleos AI vs. CBT treatment-as-usual (TAU)	First 2 months of therapy	Clinical and engagement	The AI group attended 67% more sessions (mean 5.24 vs. 3.14) and showed larger symptom reductions with depression − 34% vs. − 20% (PHQ-9), anxiety − 29% vs. − 8% (GAD-7) and no difference in satisfaction; therapists Submitted progress notes much earlier.
Danieli et al. [21]	111	Conversational AI coach for stress/anxiety management in older adults	Randomised controlled trial	Older adults with elevated stress/anxiety	Conversational-AI condition vs. minimal-intervention control	Pre-intervention, post-programme (several weeks)	Clinical	Conversational-AI group showed significantly greater reductions in perceived stress and anxiety symptoms and high acceptability among ageing adults.
McFadyen et al. [18]	540	Limbic Care AI-enabled CBT app (conversational agent + psychoeducational CBT modules)	Randomised controlled trial	Adults with elevated depression/anxiety symptoms	AI-enabled app vs. static CBT PDF workbook	6 weeks	Engagement and clinical	The AI-app group showed significantly higher engagement (more log-ins and time-on-task). Both arms showed clinical improvement; exploratory analyses suggested that participants who engaged more with AI-guided sessions experienced greater anxiety reduction.
Zhong et al. [24]	18 RCTs, 3477 participants	Various AI chatbots delivering CBT-consistent support for depression/anxiety	Systematic review and meta-analysis of randomised controlled trials	Adults with depressive and/or anxiety symptoms	Chatbots vs. various controls (wait-list, information, usual care)	Mostly short-course interventions, up to 8 weeks, with some 3-month follow-ups	Clinical	Pooled effect sizes showed significant symptom improvements: depression g − 0.26, anxiety g −0.19, with strongest effects around 8 weeks and No clear benefits at 3-month follow-up.

A second thematic area involves natural language processing (NLP) systems for detecting cognitive patterns. Jiang et al. examined the use of AI to support CBT by analysing social-media posts using the CBT-aligned ABCD model. Models such as ERNIE 3.0, PEGASUS and GPT were compared, showing ERNIE’s strength in identifying subtle cognitive patterns in smaller datasets and GPT-4’s superiority in summarisation [11]. This line of research directly informs the BECK-AI BOT’s use of a BERT-variant to detect cognitive distortions and summarise user input more accurately.

A third domain relates to chatbot-based CBT delivery, which forms the backbone of many current digital mental-health tools. Omarov et al.(2023) developed an AI chatbot therapist using AIML and RASA, offering personalised CBT through naturalistic conversations. User trials demonstrated improved mental well-being and highlighted the potential for stigma-free, accessible intervention [17]. Likewise, Danieli et al. [21] tested TEO, an AI agent used by ageing workers experiencing stress. Over eight weeks, participants receiving CBT combined with TEO experienced the greatest symptom improvement, suggesting that AI-enhanced CBT can augment traditional therapy and increase satisfaction. This supports the hybrid, clinician-supervised model embedded in the BECK-AI BOT architecture.

A fourth thematic strand highlights hybrid AI–clinician interventions and workflow enhancement. In Table 2, Sadeh-Sharvit et al. [15] evaluated Eleos Health, an AI system used to support CBT for depression and anxiety. Compared to treatment-as-usual, Eleos users showed greater symptom reduction (34% depression, 29% anxiety), greater documentation efficiency, and high satisfaction among both patients and clinicians. These findings validate the BECK-AI BOT’s clinician dashboard, risk flagging and supervision pathways. Complementing this, Frischholz et al. [14] studied a virtual CBT agent delivering cognitive restructuring across two sessions, demonstrating reductions in distress and anxiety and improvements in motivation. This underscores the importance of structured CBT exercises within digital systems, a feature central to the BECK-AI BOT design.

A fifth theme encompasses broader literature reviews exploring AI-CBT tools and therapeutic models. The review by Manole et al. [12] highlights the contribution of chatbots like Wysa, Woebot, Youper and ChatGPT in delivering CBT elements such as mindfulness and cognitive restructuring. These interventions produce reductions in anxiety and improved engagement but remain limited by short intervention periods, non-diverse samples and limited empathy, supporting the need for hybrid systems combining AI with human oversight. Recent technical reviews of Intelligent Cognitive Assistants (ICAs) highlight how user-modelling, classification-based assessment, structured dialogue trees and personalised behaviour-change mechanisms underpin the most effective text-based systems for supporting stress, anxiety and depression, providing a conceptual foundation for positioning the present chatbot within this broader landscape of ICA design [22]. This evidence directly shaped the BECK-AI BOT’s layered architecture, emphasising safety, empathy augmentation and clinician involvement. Tess, another CBT-aligned AI chatbot, is discussed in the study “Using Psychological Artificial Intelligence (Tess) to Alleviate Symptoms of Depression and Anxiety.” This work shows meaningful reductions in depression and anxiety in college students who often face barriers such as stigma or cost. Tess uses personalised feedback and integrates CBT, mindfulness and motivational interviewing techniques, reinforcing the role of AI as a supplement rather than a replacement for human therapy [13]. Finally, the review “The Use of Artificial Intelligence in Psychotherapy” provides a broad perspective on AI’s influence in mental health care, tracing the evolution from early systems like ELIZA to modern chatbots such as Woebot, Tess, and Wysa [23]. While highlighting the benefits of accessibility, diagnostic support and real-time monitoring, it also emphasises limitations in emotional depth, data ethics and therapeutic connection. These insights further support the BECK-AI BOT’s emphasis on ethical safeguards, crisis routing, data governance and human-in-the-loop oversight.

Dataset sources

This section synthesises empirical sources drawn from observational studies, experimental trials, RCTs, and engagement-based evaluations to illustrate how existing AI-enabled CBT systems have been designed, trained, and assessed. These studies form the methodological foundation for the present review and directly inform the architectural choices in the BECK-AI BOT prototype. The datasets originate from diverse populations, settings and intervention types, including mobile chatbots, virtual CBT agents, hybrid AI-therapist systems and multilingual AI counsellors. Together, these sources demonstrate the range of data types (self-report scales, conversation logs, behavioural markers, EMA-style entries, and therapist documentation) that current AI-CBT systems rely upon, thereby justifying the multi-layered data pipeline adopted in the proposed model.

Chatbot training data and evaluation approaches in existing studies

The research by Omarov et al. [17] illustrates a dataset strategy that combines everyday conversational prompts with validated clinical psychological instruments to train an AI “chatbot psychologist”. The dataset used included everyday phrases such as “How are you?” as well as more clinically significant expressions reflecting anxiety or depression. The system was also trained using established psychological tools, such as the GAD-7 and Beck Depression Inventory, enabling the recognition of appropriate symptoms. This data design is consistent with emerging ToM-ensemble architectures, where standardised clinical scales are integrated with high-frequency EMA records to train cognitive-behaviour prediction models in intelligent psychotherapy systems [10]. Thousands of conversational samples were labelled for emotion, intent and concern type, and the system was adapted for Kazakh speakers to enhance accessibility. The model achieved a 99.4% test success rate, although the authors note that greater cultural and situational diversity is required for improved generalisability [17]. These findings underscore the significance of incorporating multilingual capabilities and culturally sensitive datasets into the BECK-AI BOT methodology.

A related dataset strategy is shown in Danieli et al. [21], who evaluated TEO, an AI chatbot used by 45 Italian ageing professionals experiencing mild to moderate anxiety and occupational stress. Participants were allocated into four groups (CBT, CBT + TEO, TEO alone, and no treatment), with anxiety and stress assessed using GAD-7, PHQ-8, and PSS at baseline, mid-point, and post-intervention. User satisfaction surveys and usability feedback were also gathered. The combined CBT + TEO group demonstrated the greatest improvement in stress management, indicating that hybrid AI-human interventions can enhance therapeutic outcomes [21]. These findings support the hybrid, stepped-care approach embedded in the BECK-AI BOT.

Datasets for cognitive distortion detection and thought pathway analysis

A more linguistically structured dataset was presented in Jiang et al. [11], which examined cognitive streams in individuals with depression and suicidal ideation using 555 social media posts collected from Weibo and Reddit, translated into Chinese. Psychologists annotated the posts using the CBT-aligned ABCD framework (Activating Event, Belief, Consequence, Disputation) across 19 subcategories. The dataset included 4,742 annotated sentences, divided into training, validation, and test sets, as well as 1,643 sentence-summary pairs for therapist-usable summarisation tasks. Deep learning models, such as ERNIE 3.0, performed well for pathway classification, whereas GPT-4 occasionally produced hallucinated outputs, underscoring the need for robust safety mechanisms. These findings directly support the decision to include a Transformer-based model (MentalBERT) and safety filters in the design of the BECK-AI BOT.

RCT-Based chatbot CBT data and symptom change patterns

The Tess randomised controlled trial further demonstrates the potential of chatbots to reduce anxiety and depressive symptoms [13]. In a sample of 75 U.S. college students (74 completing), PHQ-9, GAD-7, and PANAS assessments were conducted at baseline and follow-up 2–4 weeks later. Participants were allocated to daily-check-in, bimonthly-check-in, or control (NIMH depression eBook) groups. The dataset included 14,238 chatbot message exchanges alongside user satisfaction surveys. Tess users showed significant reductions in anxiety and depression, supporting the therapeutic relevance of conversational AI [13]. These findings reinforce the clinical rationale behind implementing CBT-aligned dialogue flows in BECK-AI BOT.

Short-Term observational trials demonstrating anxiety reduction

The data in Manole et al. [12], summarised in Table 1, adds evidence on short-term AI-CBT interventions. In a two-phase observational trial (N = 50), volunteers with mild–moderate anxiety submitted daily anxiety and sleep ratings, as well as interaction patterns with a ChatGPT-based CBT chatbot across two seven-day intervals, two months apart. Significant anxiety reductions were observed (21.15% in Phase 1; 20.42% in Phase 2), with qualitative feedback emphasising usability and empathy limitations. These datasets reveal both benefits and gaps in current chatbot implementations, guiding improvements in empathy modelling, personalisation and safety for BECK-AI BOT.

Large-Scale engagement datasets and AI-Assisted clinical care

In McFadyen et al. [18], a large RCT of 540 adults compared a static PDF workbook with an AI-facilitated CBT app, Limbic Care [18]. Weekly GAD-7 and PHQ-9 assessments were collected over a six-week period, along with demographic variables and usage analytics. Although symptom reductions were similar across groups, the AI-assisted group demonstrated markedly higher engagement (2.4× more frequent use; 3.8× longer duration). Users who employed guided sessions achieved greater improvements in well-being (≈ 29.81%), and no additional safety concerns were reported [18]. These results support the use of engagement optimisation algorithms such as the TCN in BECK-AI BOT.

Hybrid Human-AI clinical Documentation and supervision datasets

The Sadeh-Sharvit et al. [15] study evaluated Eleos Health’s AI-supported CBT enhancement platform in a naturalistic outpatient setting with 47 adults diagnosed with anxiety or depression. Baseline and follow-up GAD-7/PHQ-9 scores, session transcripts, therapist documentation time and usability feedback formed the dataset. AI-CBT participants demonstrated improvements in treatment efficiency and adherence to CBT principles compared with TAU [15]. These findings justify the inclusion of a clinician dashboard and AI-supported documentation tools in BECK-AI BOT.

Experimental virtual CBT agent datasets

Frischholz et al. [14] utilised data from a two-session CBT intervention delivered by a virtual agent to 35 Japanese adults, with assessments including the QIDS-SR, STAI, and K6 administered before and after each session. The Cognitive Change Immediate and Sustained scales captured thought restructuring outcomes, while motivation was tracked using a SOCRATES-derived measure. The results demonstrated reductions in depressive and anxiety symptoms and increases in cognitive restructuring, supporting the feasibility of AI-mediated CBT [14].

High-Stress populations and crisis context datasets

Finally, Spytska et al. [23] examined AI-supported therapy among 104 Ukrainian women living in active conflict zones. Baseline and post-intervention anxiety assessments (Beck Anxiety Inventory; Hamilton Anxiety Rating Scale) were collected, alongside demographic data, psychiatric history and eligibility screening. During an eight-week intervention, the experimental group used an AI chatbot daily, with conversation logs, sentiment analysis and adaptive therapeutic responses recorded. The control group received psychotherapy from licensed clinicians. All data were anonymised, with robust safety protocols due to the high-risk setting [23]. These findings directly inform the crisis-management and safety-escalation components in BECK-AI BOT.

The studies reviewed collectively show that AI-supported CBT systems can improve symptoms, enhance engagement and support clinical decision-making across diverse populations. These findings informed the design of the BECK-AI BOT, which combines structured assessment, advanced NLP, temporal forecasting, and safety governance into a single, integrated framework. Although this work presents no new clinical trial data, it establishes a clear foundation for future empirical testing and the ethical deployment of AI-assisted mental health interventions.

Proposed methodology

The proposed working model given in Fig. 1 is designed to provide effective therapeutic intervention integrated with AI-driven conversational therapy, Cognitive Behavioral Therapy (CBT), and real-time monitoring in association with professional expertise to deliver personalized mental health support. Manole et al., in his study of AI chatbot intervention, mentioned in Table 3, laid the foundation for the effectiveness of using ChatGPT in the treatment of anxiety. The study was conducted in two phases and showed a significant reduction in anxiety-related disorders and stress disorders [12]. Also, in the survey conducted by Zheng and Ye, they predicted how effectively the user responds to CBT and provided the treatment protocol after going through multiple review patterns and cognitive diagnoses [26]. The proposed model in the study integrates both an AI Chatbot intervention using conversation using advanced Natural Language Processing (NLP) like RASA and predictive analytics using a Deep learning model. A hypothetical version of a chatbot was designed and named after Dr. Aaron T. Beck, who is recognized as the father of Cognitive Behaviour Therapy (CBT), Since he developed this treatment in early 1960’s and 1970’s, It has created a great impact and served as an effective treatment strategy for mental disorder, psychological disorders and medical conditions [29] The BECK-AI BOT (https://beck-ai-bot.b12sites.com/index) is a prototype that can be further developed into a working model by integrating all those aspects mentioned in the proposed methodology. The app interface was designed in a user-friendly manner, and its future working will provide better access to both patients and clinicians.

Fig. 1.

The figure represents the proposed model of the study; here, the multi-stage assessment is undertaken through General Anxiety Scale-7 (GAD-7), Patient Health Questionnaire-9 (PHQ-9), Beck Anxiety Inventory (BAI), and Ecological Momentary Assessment (EMA). The open AI used in the study is Reasonable Artificial Service Agent (RASA), and the dual-deep learning models used were Transformed – Natural Language Processing (BERT variant) and Temporal Convolutional Network (TCN).

Table 3.

Technical machine-learning studies related to CBT and mental-health prediction

Reference	Model	Task	Population	Outcome	Key Technical outcome
[25]	MentalBERT and MentalRoBERTa (domain-adapted BERT/RoBERTa)	Mental-health text classification (depression, stress, suicidal ideation, etc.)	Multiple Reddit and benchmark datasets for mental-health detection	ML accuracy	Domain-specific models improved recall and F1 on several mental-health detection benchmarks; MentalRoBERTa often achieved the best F1 across tasks.
[26]	Deep-learning correlation prediction model (multi-objective evolutionary algorithm)	Predict CBT treatment effects on adolescent social anxiety	Adolescents with social anxiety: CBT evaluation system with risk and protective factors	ML accuracy	Deep-learning model showed higher prediction accuracy and reduced model complexity compared with traditional approaches; able to predict CBT effects and alleviate social anxiety in simulation.
[27]	SVM classifier based on fMRI (dACC–amygdala BOLD responses)	Predict long-term iCBT response (responder vs. non-responder)	26 adults with social anxiety disorder receiving iCBT	ML accuracy (prognostic)	SVM using dACC–amygdala activation predicted 1-year response with 92% accuracy; neural patterns differentiated long-term responders from non-responders.
[16]	Transformer-based deep neural networks predicting Cognitive Therapy Rating Scale (CTRS) codes	Automated assessment of CBT fidelity from session audio/transcripts	2,494 recorded CBT sessions with human CTRS ratings	ML accuracy/fidelity metrics	AI-generated CTRS scores reached ≥ 80% of human reliability for almost all CTRS items and 100% for the total score; this demonstrates the feasibility of automated CBT quality scoring.
[28]	Generic CNN architectures for text classification	Classify mental-health related text (e.g., distress categories)	Benchmark mental health corporations	ML accuracy	CNN models achieved competitive or superior accuracy and F1 compared with baseline models for mental-health text classification.

The model in Fig. 1 follows a layered architecture. The BECK-AI BOT introduced in this research employs a tiered AI-driven behavioural therapy (CBT) framework that combines conversational assistance, multimodal evaluation, predictive analytics, and organized safety management. Since this study does not include subjects, the methods section details the theoretical and technical design of the system, along with the intended offline assessment strategy. Users enter the platform via a secure web portal that features two-factor authentication, during which they are prompted to provide demographic data and give their informed consent. Every piece of health data is handled adhering to privacy-by-design standards employing AES-256 encryption for stored data, TLS 1.3 for secure transmission, and role-based permissions for healthcare providers [30,31,32]. After signing up, the platform conducts a psychological evaluation using established clinical tools, such as the GAD-7, PHQ-9, and BAI, supplemented by Ecological Momentary Assessment (EMA) notifications sent one to three times daily, which provides enhanced monitoring [18].

Such prompt-based sampling aligns with established methodological approaches in smartphone mental health research, where high-frequency EMA and just-in-time delivery are used to capture dynamic changes in affect and guide moment-appropriate intervention strategies [5]. These EMA records track moment-to-moment changes in mood, stress levels, sleep quality and behaviour, producing data utilised for predictive models. Natural language comprehension is managed by a RASA-driven pipeline that executes detection and entity identification [17]. User inputs are sorted into relevant classes, such as stress expression, negative automatic thoughts, avoidance actions, problem explanations, relapse-associated terminology and crisis or self-harm signals. Extracted entities highlight emotions, triggers, indicators of severity and temporal references serving as the input for reasoning and intervention systems. A refined psychological analysis is offered by MentalBERT, a transformer model tailored to the domain, pretrained on extensive mental health text collections and further fine-tuned using publicly accessible datasets like CLPsych and SMHD [25]. MentalBERT detects nuances, cognitive distortions (such as catastrophizing, overgeneralizing) and thinking patterns that guide the choice of CBT methods [11].

The system’s second fundamental element is a Temporal Convolutional Network (TCN) created to capture fluctuations in EMA data. Utilising a fourteen-day timeframe and dilated causal convolutions, the TCN forecasts next-day symptom intensity, probability of disengagement and short-term relapse risk [28,18]. These forecasts enable the adjustment of interventions, with the system providing cognitive restructuring tasks, behavioural activation methods, psychoeducation, or guided relaxation tailored to each user’s changing requirements. When ongoing catastrophising or rising distress is identified, the model prioritises restructuring elements and grounding strategies; when disengagement is anticipated, it intensifies behavioural cues. Safety is integrated at every layer of the system’s design. A crisis-detection algorithm, adapted from MentalBERT, detects self-harm expressions by applying a probability cutoff of 0.85 [25,11]. This design aligns with contemporary ToM-simulating cognitive architectures that integrate user modelling, machine-learning classifiers, and structured psychological ontologies to forecast symptom trajectories and personalise interventions dynamically, as demonstrated in recent computational psychotherapy frameworks [4]. Utterances deemed high-risk initiate an instant crisis response, where the system delivers comforting grounding messages, limits urgent conversation and initiates location-based routing to appropriate emergency numbers, like 112, 999 or 988 [33]. At the time, a clinical alert is sent via the dashboard, and the session is locked until a professional assessment is finished. This human-, in-the-loop governance framework lowers the chance of system malfunction and ensures clinical supervision in critical situations [15,16].

The clinician dashboard presents shifts in assessment scores emphasizes high-risk statements showcases TCN-generated decline trajectories and provides summaries of cognitive distortions detected by MentalBERT. Clinicians have the ability to examine transcripts modify treatment advice and oversee high-risk cases ensuring consistency with stepped-care frameworks [15,16]. Ethical measures involve pseudonymizing user data, recording automated decisions for auditing purposes and conducting bias audits across groups to detect discrepancies, in intent recognition or crisis identification [32,31].

In contrast to AI-driven CBT platforms like Woebot, Wysa, Tess, Limbic and Eleos the BECK-AI BOT presents multiple enhancements. Integrating RASA with MentalBERT enables the system to better identify distortions and engage in more adaptable conversations compared to rule-based chatbots [17,12,13]. The choice of a RASA-based dialogue manager is consistent with findings from state-of-the-art ICA architectures, which show that hybrid dialogue policies combine intent classification with rule-based planning, which will achieve greater conversational stability and therapeutic alignment than fully rule-based systems alone [22]. The incorporation of TCN forecasting allows for long-term personalisation not in existing solutions and EMA-based adjustments synchronize the timing and intensity of interventions, with the users immediate emotional condition [18]. The inclusion of a clinician dashboard, multi-tiered safety escalate routes and privacy-by-design structures further distinguishes the model as a next-generation hybrid system that unifies conversational CBT, predictive analytics, crisis governance and professional oversight [23].

Safety, ethics and clinical governance

The BECK-AI BOT is designed according to adhering to safety protocols and clinical governance principles, which are appropriate for present digital mental health systems, Note that no human participant data were collected for this study, and the prototype has not been deployed in a real-world setting.

While a user enters the portal, informed consent is obtained, which is consistent with the APA telepsychology guidelines [30] and WHO recommendations for health interventions [31]. Data minimisation principles ensure that only essential information, such as demographics, symptom scores and EMA entries, is collected, with no use of identifiable text for model training. It is in accordance with OECD Artificial Intelligence principles [32]. All data are encrypted in transit using TLS 1.3 and at rest using AES-256, and access is restricted through role-based permissions for clinicians. As discussed above in the methodology, risk monitoring is layered and continuously screens for harms, suicidal intent and hopelessness. Continuous risk monitoring will identify severe EMA scores and emotional deterioration over weeks and will lead to activation of the crisis protocol. Responses will be generated accordingly, with severe attention and emergency service numbers of that specific geographic location provided. A corresponding alert is sent to the clinician dashboard, creating a clear handoff. Service-level expectations require clinicians to respond to red-flag alerts within a defined time frame; if no clinician is available, the system directs the user exclusively to human emergency resources.

Transparency, accountability and fairness considerations draw from open-access AI ethics scholarship, including Weidinger et al.’s taxonomy of risks for large language models (2022) [33]. In line with these frameworks, BECK-AI BOT maintains a strict duty-of-care boundary by delivering only low-intensity CBT-based support while ensuring that diagnostic decision-making and crisis management remain under human clinical supervision. No human data were collected for this manuscript, and all technical descriptions represent the governance processes intended for future evaluation. From a regulatory perspective, the system is intended to align with Software as a Medical Device (SaMD) criteria as defined by the FDA, MHRA and EU AI Act. The post-market surveillance plan includes periodic safety audits, error-rate monitoring, regular retraining with bias-audited datasets, user feedback collection and transparent documentation of model updates. These measures ensure that the system evolves within an ethical, clinically governed, and legally compliant framework.

Discussion

Cognitive behavioural therapy (CBT) is a kind of psychological therapy that is effective for many kinds of psychological interventions, such as depression, alcohol and drug use problems, anxiety disorders, eating disorders, and severe mental illness. CBT significantly enhances functioning and quality of life, according to numerous studies, as shown in Fig. 2. Numerous studies have demonstrated that CBT is as effective as, or even more effective than, other forms of mental treatment or psychiatric medications. Cognitive-behavioural therapy (CBT) continues to be a fundamental pillar in managing mental health disorders, with contemporary studies highlighting its effectiveness, versatility, and lasting advantages. A research paper in The Lancet Psychiatry showed that CBT for treatment-resistant depression resulted in sustained gains over 46 months, with 43% of the participants showing ≥ 50% symptom reduction versus 27% in usual care [34]. These findings highlight CBT’s cost-effectiveness in real healthcare settings and close long-standing evidence gaps regarding its long-term effects, particularly in cases of complex depression. These findings support CBT’s position not just as a short-term treatment but as a skills-based therapy with long-standing effects on depressive relapse and the value of life [34]. CBT has applicability in comorbid physical and psychological disorders. CBT was shown to decrease depressive symptoms (effect sizes d = 1.31–0.18 post-treatment) and anxiety (d = 1.08–0.19) significantly in a 2023 systematic review of 1,661 individuals with chronic pain and psychological distress [35]. There was no sustained improvement in pain intensity or catastrophizing, suggesting that psychological rather than sensory processes mediate the effects of CBT [36]. This demarcation highlights the necessity of integrated treatment models that merge CBT with somatic therapies to achieve comprehensive chronic pain management.

Fig. 2.

Diagram illustrating the strengths and limitations of CBT and its engagement in psychological intervention

The emergence of digital CBT modalities driven by the COVID-19 pandemic has improved access without a decrease in efficacy. Bibliometric reviews indicate a 12.67% yearly increase in CBT research since 1979, with new trends in mobile health and internet-delivered interventions [20]. Online therapy and app-dispensed modules show similar effects to face-to-face therapy for conditions such as generalized anxiety and mild to moderate depression [36]. The diversity of study populations and designs, however, suggests that a particular implementation is required. For example, a review conducted in 2024 noted that while digital cognitive behavioral therapy reduces barriers to care, its cost-effectiveness varies by population and depends on variables like the severity of symptoms and technological literacy [20]. Emerging developments have broadened CBT’s use. Written Exposure Therapy (WET), a short CBT subtype for PTSD, holds promise in alleviating avoidance behaviours using less-than-standard procedures. For youth with OCD, modular forms of CBT that integrate exposure-response prevention with family-based strategies enhance adherence and outcomes among younger groups. In the same way, CBT interventions for ADHD in teenagers now aim at functional areas such as time management and emotion regulation, with meta-analyses indicating moderate-to-large effect sizes in academic and social functioning. Despite its adaptability, CBT has drawbacks. There is still a lack of extended data for non-depressive disorders after 4–5 years [36]. Cultural and societal barriers also limit global implementation; a 2024 Saudi Arabian survey found that a lack of training and public awareness is a significant bottleneck [20]. CBT remains a developing evidence-based intervention with solid applications in both mental and physical well-being arenas. Its interoperability with digital platforms and individualized methods highlights its flexibility, though scalability must address systemic mental healthcare access inequities.

The established strengths of CBT highlight the need for AI systems that can match its flexibility and skill-building potential. Overall, the reviewed literature shows that while AI-assisted CBT systems offer meaningful benefits, several gaps remain that justify the design of the BECK-AI BOT. Personalisation remains limited in many AI tools, as evidenced by studies by Manole et al. (2024), Omarov et al. (2023), and Jiang et al. (2024), which reveal that chatbots struggle with emotion detection and NLP models encounter issues such as confusion or imbalanced category detection [12,17, 11]. In contrast to established conversational agents such as Woebot and Wysa, the authors of recent computational psychotherapy systems report that ToM-driven modelling can capture a wider range of cognitive categories and yield more accurate symptom trajectory forecasts in their cohorts [4], highlighting a direction that supports the forecasting component of the BECK-AI BOT. McFadyen et al. [18] demonstrated that AI-guided CBT apps substantially increase engagement, aligning with the historical trend of digital CBT growth noted in bibliometric studies. The BECK-AI BOT builds on this by integrating adaptive engagement features, dynamic EMA prompting, and predictive modelling to support long-term use. Ethically, the evidence highlights concerns about empathy limitations, crisis safety, and potential bias, reinforcing the need for strong governance frameworks in AI systems. Clinical integration findings further indicate that AI works best as a complement to therapists [16,15], not a replacement, as shown in hybrid models such as Eleos Health and LyssnCBT. Taken together, these findings suggest that future AI-CBT systems must focus on deeper personalisation, culturally informed datasets, rigorous safety protocols, and hybrid human–AI supervision that are central to the BECK-AI BOT’s proposed model.

Conclusion and future scope

The proposed methodology and model, as shown in Fig. 3, consist of AI-driven assisted therapy with interactive conversational analysis and predictive analysis using the RASA open AI system, which provides tailored intervention protocols and treatment strategies. The potential of AI chatbots in Cognitive Behavioural Therapy (CBT) in the future is very promising, fueled by advances in artificial intelligence and the growing need for affordable mental healthcare. These findings align with persuasive-technology principles, which stress that effective mental-health interventions for low-resource settings must prioritise usability, low digital burden and affordability to ensure equitable access across diverse populations [7]. AI chatbots such as Woebot, Wysa, and Youper have shown impressive symptom reduction in depression and anxiety, with high user engagement and satisfaction [37]. These chatbots provide individualised interventions, such as cognitive restructuring, behavioural activation, and mindfulness exercises, while facilitating remote monitoring of mental health improvement. New technologies like large language models (LLMs) and wearable technologies are improving the capabilities of chatbots by applying predictive analytics and real-time feedback during therapy. Overall, the targets of future studies are long-term effectiveness, scalability, and integration into mental health systems. Issues like emotional sensitivity, cultural responsiveness, and ethical concerns are still prioritized for improvement. No clinical evaluation of the BECK-AI BOT was conducted in this study. The improvements discussed reflect outcomes reported in prior research on comparable AI-supported CBT tools. Future work will involve pilot testing and formal evaluation of the proposed system. It is worth noting that AI chatbots will augment human therapy, rather than replace it.

Fig. 3.

Screenshots of the proposed WebApp interface of the AI-driven mental chatbot

Author contributions

AS: Writing - Original draft preparation, Investigation, Methodology. MS: Data curation, Visualization, MR: Conceptualization, Validation, Writing- Reviewing and Editing. NK: Discussion, Reviewing, Formatting. CSE: Methodology, Investigation. KS: Supervision, Final validation, Proofreading and Editing, Investigation and Planning.

Funding

Not applicable.

Data availability

Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.

Declarations

Ethical approval

Not applicable.

Consent to participate

Not applicable.

Consent to publish

Not applicable.

Clinical trial number

Not applicable.

Competing interests

The authors declare no competing interests.