Charting the evolution of artificial intelligence mental health chatbots from rule‐based systems to large language models: a systematic review

Abstract

The rapid evolution of artificial intelligence (AI) chatbots in mental health care presents a fragmented landscape with variable clinical evidence and evaluation rigor. This systematic review of 160 studies (2020‐2024) classifies chatbot architectures – rule‐based, machine learning‐based, and large language model (LLM)‐based – and proposes a three‐tier evaluation framework: foundational bench testing (technical validation), pilot feasibility testing (user engagement), and clinical efficacy testing (symptom reduction). While rule‐based systems dominated until 2023, LLM‐based chatbots surged to 45% of new studies in 2024. However, only 16% of LLM studies underwent clinical efficacy testing, with most (77%) still in early validation. Overall, only 47% of studies focused on clinical efficacy testing, exposing a critical gap in robust validation of therapeutic benefit. Discrepancies emerged between marketed claims (“AI‐powered”) and actual AI architectures, with many interventions relying on simple rule‐based scripts. LLM‐based chatbots are increasingly studied for emotional support and psychoeducation, yet they pose unique ethical concerns, including incorrect responses, privacy risks, and unverified therapeutic effects. Despite their generative capabilities, LLMs remain largely untested in high‐stakes mental health contexts. This paper emphasizes the need for standardized evaluation and benchmarking aligned with medical AI certification to ensure safe, transparent and ethical deployment. The proposed framework enables clearer distinctions between technical novelty and clinical efficacy, offering clinicians, researchers and regulators ordered steps to guide future standards and benchmarks. To ensure that AI chatbots enhance mental health care, future research must prioritize rigorous clinical efficacy trials, transparent architecture reporting, and evaluations that reflect real‐world impact rather than the well‐known potential.

Mental disorders remain a major contributor to the global burden of disease. An estimated 970 million individuals live with mental health or substance use disorders worldwide, with depression and anxiety among the leading causes of disability‐adjusted life years lost ¹ , ² .

Despite the increasing recognition of mental health as a critical component of public health, care systems remain underfunded and overwhelmed ³ , with fewer than five mental health professionals available per 100,000 people globally ⁴ . This unmet need is more pronounced in low‐ and middle‐income countries, where more than 75% of individuals with mental health conditions fail to receive treatment ⁵ . The need for scalable, accessible solutions has driven interest in digital interventions, particularly conversational agents or chatbots, as potential tools to support mental health care by way of screening, psychoeducation, and therapy augmentation ⁶ .

While the rapid growth of artificial intelligence (AI)‐driven chatbots offers new possibilities for mental health applications, their potential remains unclear. Many clinicians, policy makers and researchers are uncertain about whether these tools are appropriate for clinical deployment. Much of this uncertainty arises from the heterogeneous nature of chatbot technologies, which range from basic rule‐based systems to advanced large language models (LLMs). While rule‐based chatbots rely on pre‐programmed scripts or decision trees, LLMs leverage deep neural networks trained on vast datasets to produce more versatile, human‐like conversational capabilities.

The conflation of these technologies has led to several challenges. While most health care chatbots remain rule‐based, companies often continue to market them as “AI” or even “LLM‐driven”. This creates misconceptions about their sophistication and reliability, as these systems differ substantially in their capabilities and limitations. Additionally, exaggerated claims about LLMs, such as their ability to pass professional exams or display empathy surpassing that of human clinicians, have blurred the distinction between experimental results and real‐world applicability ⁷ , ⁸ . For instance, passing a written test or simulating empathetic dialogue in controlled conditions does not necessarily equate to making accurate clinical diagnoses or supporting patients in complex and dynamic clinical scenarios.

The current state of the AI chatbot literature reflects these challenges. Meta‐analyses and reviews often conflate simple feasibility testing results with more complex clinically focused research. In this paper, we applied a staged approach to translational research and clinical development that mirrors the progression from basic science/preclinical research to early phase human testing (phase 1 or 2 clinical trials) to clinical efficacy testing (phase 3 clinical trials). This scheme organizes chatbot studies into three tiers: foundational bench testing for technical feasibility, pilot feasibility testing for assessment of usability and acceptability in humans, and clinical efficacy testing. These categories reflect increasing levels of clinical applicability, providing a structured approach to understanding the field's progress. This approach contextualizes the current discussion around AI certification with a staged structure ⁹ , and provides a useful scaffold for understanding prior studies and developing future work.

The evolution of AI chatbots reflects decades of advancements in computational approaches to human language, which we categorize into three paradigms: rule‐based systems, non‐LLM machine learning models, and LLM‐based systems. This tripartite framework is designed to clarify distinctions between these paradigms, particularly as the term “AI” has historically encompassed a wide range of technologies, from early deterministic systems to modern probabilistic models.

Rule‐based systems were the earliest form of conversational AI. ELIZA, developed in the 1960s, exemplified these systems by simulating a Rogerian psychotherapist through simple pattern‐matching and substitution rules ¹⁰ . Currently popular mental health chatbots, such as Woebot, have been primarily rule‐based systems, underscoring the enduring effectiveness of this approach ¹¹ . Most chatbots continue to be employed in narrowly defined tasks, such as structured screening tools or symptom checkers ¹² , where deterministic outputs remain sufficient. However, their inability to adapt to novel inputs or provide personalized responses has constrained their utility where dynamic and context‐sensitive interactions are critical ¹³ . Despite limitations, rule‐based systems laid the groundwork for subsequent innovations by demonstrating the feasibility of automated dialogue.

The transition from rule‐based systems to machine learning introduced greater adaptability and probabilistic reasoning into chatbot designs. Machine learning refers to algorithms that identify patterns in data, enabling models to generalize beyond explicit rules and make predictions based on statistical likelihoods ¹⁴ , ¹⁵ . Non‐LLM machine learning‐based chatbots marked a pivotal shift by moving beyond scripted interactions. These systems incorporated natural language processing techniques, such as sentiment analysis and intent recognition, to infer user emotions and tailor responses to context ¹⁶ . For example, conversational agents such as Wysa employ a combination of machine learning algorithms and rule‐based scripts to deliver cognitive‐behavioral therapy (CBT) interventions, demonstrating efficacy in reducing symptoms of depression and anxiety in pilot studies ¹⁷ , ¹⁸ .

While non‐LLM machine learning models expanded the scope of chatbot applications, they faced inherent challenges. Their performance often depends on domain‐specific training data, which limits generalizability across diverse conversational scenarios. Additionally, these systems struggle with generating coherent and human‐like language, as they were typically designed to classify or process inputs rather than produce contextually appropriate outputs.

LLMs represent a paradigm shift in AI chatbots, driven by their ability to generate human‐like language with fluency and contextual awareness. Built on Transformer architectures, LLMs such as OpenAI's GPT series and Meta's Llama models leverage self‐attention mechanisms to understand relationships between words and concepts across extensive text passages ¹⁹ , ²⁰ . This architecture enables LLMs to maintain coherence within complex, multi‐turn dialogues in ways that previous approaches could not achieve ²¹ . Unlike earlier machine learning‐based systems that primarily focused on classification or limited response selection, LLMs are fundamentally generative in nature – they create novel text rather than selecting from predefined responses.

LLMs’ advanced linguistic capabilities have sparked explosive interest in their mental health applications ²² , offering the potential to enhance psychoeducation, triage, and supportive interactions. However, the transformative potential of LLMs is accompanied by significant challenges. Their reliance on vast, uncurated datasets introduces risks such as bias, misinformation, and the generation of fabricated or harmful content ²³ . In psychiatry, where the consequences of errors can be severe, these limitations raise ethical concerns about reliability and safety.

METHODS

We classified chatbots into three systems:

• Rule‐based systems. These rely on deterministic scripts (e.g., rule‐based conversation systems, simple decision trees), with no data‐driven learning. They are ideal for structured, low‐risk tasks (e.g., symptom checklists) where predictability ensures safety. However, their rigidity limits their utility in dynamic therapeutic contexts.

• Machine learning‐based systems. These include traditional machine learning (e.g., support vector machine, SVM) and non‐generative deep learning (e.g., recurrent neural networks, RNN; long short‐term memory, LSTM; and bidirectional encoder representations from transformers, BERT). While RNN/LSTM and traditional machine learning differ technically (e.g., sequential vs. static data processing), both lack natural language fluency. Grouping these under “machine learning‐based systems” reflects their shared limitation in mental health: adaptability without generative capacity.

• LLM‐based systems. These leverage generative models trained on vast text corpora to produce human‐like dialogue. This category includes multimodal models that can process images, audio or other modalities in addition to text, as long as they maintain the core LLM architecture for language generation.

We used a tiered framework that categorizes studies by their evaluation rigor, akin to the translational pipeline from technical validation to real‐world clinical impact:

• T1. Foundational bench testing. This focuses on technical validation in controlled settings (e.g., scripted scenarios, expert assessments) to ensure that chatbots meet baseline functional and safety standards. For mental health, this stage is critical to verify adherence to clinical guidelines (e.g., suicide risk protocols) before human interaction.

• T2. Pilot feasibility testing. This assesses usability and acceptability with human participants (e.g., patients, clinicians) over short‐term interactions. While it provides insights into engagement, it often overlooks sustained therapeutic outcomes – a gap particularly problematic in mental health, where longitudinal efficacy is paramount.

• T3. Clinical efficacy testing. This measures clinically meaningful outcomes (e.g., symptom reduction via validated rating scales) over extended periods. It is essential for mental health chatbots, as transient usability gains (T2) do not necessarily equate to therapeutic benefit.

This framework ensures that mental health interventions undergo rigorous validation before clinical deployment. A chatbot that performs well in scripted tests (T1) may still fail in real‐world empathy or crisis management, while short‐term usability (T2) does not guarantee long‐term adherence or relapse prevention. By stratifying evidence into three tiers, the classification enables clinicians and regulators to distinguish technically functional tools from those with proven clinical impact ²³ .

We systematically reviewed mental health chatbot studies published from January 1, 2020 to January 1, 2025 across PubMed, APA PsycNet, Scopus and Web of Science, following PRISMA 2020 guidelines ²⁴ . Search strings were adapted from prior scoping reviews and optimized for lexical coverage (see supplementary information). Additional records were identified through manual searches of Google Scholar and major AI conference proceedings.

To ensure relevance, only studies evaluating chatbots within a mental health care context were included. Papers focused on psycholinguistics, psychosocial demographics, or predictive models without conversational interfaces were excluded. Only full, peer‐reviewed papers written in English were considered. Reviews, meta‐analyses and retracted papers were excluded. Eligible studies were required to include an actual evaluation of chatbot performance, excluding protocol descriptions.

The screening and annotation process was a collaborative effort involving the entire research team. Each study was randomly assigned to at least two reviewers, who independently evaluated its eligibility based on predefined inclusion criteria. Any disagreements between reviewers were resolved through group discussions to ensure consistency.

Figure 1 presents the PRISMA flow diagram illustrating the study selection process. A total of 1,727 records were identified through database and manual searches, including 620 from PubMed, 18 from APA PsycNet, 419 from Scopus, 480 from Web of Science, 131 from Google Scholar, and 59 from major AI conferences. After removing 790 duplicates, 937 unique records were screened. Of these, 734 studies were excluded based on title and abstract screening, due to irrelevance or failure to meet the inclusion criteria. Following this, 203 reports were sought for full‐text retrieval, but four could not be retrieved. The remaining 199 full‐text reports were assessed for eligibility, leading to the exclusion of eight studies that lacked chatbot evaluation, 21 that were duplicates published under different titles, and ten whose models did not meet our definition of LLMs. Ultimately, 160 studies met the inclusion criteria and were included in the systematic review ²⁵ , ²⁶ , ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ , ³² , ³³ , ³⁴ , ³⁵ , ³⁶ , ³⁷ , ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴² , ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ , ⁵² , ⁵³ , ⁵⁴ , ⁵⁵ , ⁵⁶ , ⁵⁷ , ⁵⁸ , ⁵⁹ , ⁶⁰ , ⁶¹ , ⁶² , ⁶³ , ⁶⁴ , ⁶⁵ , ⁶⁶ , ⁶⁷ , ⁶⁸ , ⁶⁹ , ⁷⁰ , ⁷¹ , ⁷² , ⁷³ , ⁷⁴ , ⁷⁵ , ⁷⁶ , ⁷⁷ , ⁷⁸ , ⁷⁹ , ⁸⁰ , ⁸¹ , ⁸² , ⁸³ , ⁸⁴ , ⁸⁵ , ⁸⁶ , ⁸⁷ , ⁸⁸ , ⁸⁹ , ⁹⁰ , ⁹¹ , ⁹² , ⁹³ , ⁹⁴ , ⁹⁵ , ⁹⁶ , ⁹⁷ , ⁹⁸ , ⁹⁹ , ¹⁰⁰ , ¹⁰¹ , ¹⁰² , ¹⁰³ , ¹⁰⁴ , ¹⁰⁵ , ¹⁰⁶ , ¹⁰⁷ , ¹⁰⁸ , ¹⁰⁹ , ¹¹⁰ , ¹¹¹ , ¹¹² , ¹¹³ , ¹¹⁴ , ¹¹⁵ , ¹¹⁶ , ¹¹⁷ , ¹¹⁸ , ¹¹⁹ , ¹²⁰ , ¹²¹ , ¹²² , ¹²³ , ¹²⁴ , ¹²⁵ , ¹²⁶ , ¹²⁷ , ¹²⁸ , ¹²⁹ , ¹³⁰ , ¹³¹ , ¹³² , ¹³³ , ¹³⁴ , ¹³⁵ , ¹³⁶ , ¹³⁷ , ¹³⁸ , ¹³⁹ , ¹⁴⁰ , ¹⁴¹ , ¹⁴² , ¹⁴³ , ¹⁴⁴ , ¹⁴⁵ , ¹⁴⁶ , ¹⁴⁷ , ¹⁴⁸ , ¹⁴⁹ , ¹⁵⁰ , ¹⁵¹ , ¹⁵² , ¹⁵³ , ¹⁵⁴ , ¹⁵⁵ , ¹⁵⁶ , ¹⁵⁷ , ¹⁵⁸ , ¹⁵⁹ , ¹⁶⁰ , ¹⁶¹ , ¹⁶² , ¹⁶³ , ¹⁶⁴ , ¹⁶⁵ , ¹⁶⁶ , ¹⁶⁷ , ¹⁶⁸ , ¹⁶⁹ , ¹⁷⁰ , ¹⁷¹ , ¹⁷² , ¹⁷³ , ¹⁷⁴ , ¹⁷⁵ , ¹⁷⁶ , ¹⁷⁷ , ¹⁷⁸ , ¹⁷⁹ , ¹⁸⁰ , ¹⁸¹ , ¹⁸² , ¹⁸³ , ¹⁸⁴ .

Figure 1.

PRISMA flow diagram of study selection process. LLMs – large language models

To ensure consistency and depth in data extraction, senior team members (YH and JT) implemented a structured annotation protocol (see supplementary information), formalized through team training. Key elements included chatbot architecture, evaluation methodology, target conditions, functional purpose, and outcome measures. Each study was also annotated with type‐specific information based on its evaluation tier (T1, T2 or T3), capturing evaluator characteristics, usage duration, and relevant clinical instruments. Missing or non‐applicable data were systematically flagged to support transparent synthesis.

To analyze the 160 annotated studies, we implemented a structured multi‐step methodology to transform raw annotations into standardized themes. In cases where studies were classified into multiple categories, each classification was weighted proportionally (e.g., a study targeting both depression and anxiety would be counted twice, each with weighting 0.5). Reported subtotals and percentages were rounded to the nearest whole number, which may result in apparent summation discrepancies. Computational tools supported initial categorization, with all results refined and validated by domain experts to ensure fidelity and interpretability (see also supplementary information).

RESULTS

Evolution of chatbot architectures

Research interest in mental health chatbots increased substantially over the review period, with the annual number of studies quadrupling from 14 in 2020 to 56 in 2024. Coinciding with this growth, the underlying chatbot architectures underwent a significant transformation (see Figure 2).

Figure 2.

Evolution of chatbot architectures studied from 2020 to 2024. Percentages indicate the proportion of studies utilizing each architecture type within a given year. The number above each bar indicates the total number of studies for that year. ML – machine learning, LLM – large language model.

Initial research in 2020 (n=14) ²⁵ , ²⁶ , ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ , ³² , ³³ , ³⁴ , ³⁵ , ³⁶ , ³⁷ , ³⁸ focused exclusively on rule‐based systems (100%). The landscape diversified from 2021 (n=28) ³⁹ , ⁴⁰ , ⁴¹ , ⁴² , ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ , ⁵² , ⁵³ , ⁵⁴ , ⁵⁵ , ⁵⁶ , ⁵⁷ , ⁵⁸ , ⁵⁹ , ⁶⁰ , ⁶¹ , ⁶² , ⁶³ , ⁶⁴ , ⁶⁵ , ⁶⁶ , with the emergence of machine learning‐based (21%) and the first LLM‐based studies (11%), although rule‐based systems remained dominant (68%). Machine learning‐based approaches peaked in 2022 (40% of 25 studies ⁶⁷ , ⁶⁸ , ⁶⁹ , ⁷⁰ , ⁷¹ , ⁷² , ⁷³ , ⁷⁴ , ⁷⁵ , ⁷⁶ , ⁷⁷ , ⁷⁸ , ⁷⁹ , ⁸⁰ , ⁸¹ , ⁸² , ⁸³ , ⁸⁴ , ⁸⁵ , ⁸⁶ , ⁸⁷ , ⁸⁸ , ⁸⁹ , ⁹⁰ , ⁹¹ ), before stabilizing as a smaller component (14% in 2024).

LLM‐based architectures, after comprising 16% of studies in 2022 and 19% in 2023 (n=37) ⁹² , ⁹³ , ⁹⁴ , ⁹⁵ , ⁹⁶ , ⁹⁷ , ⁹⁸ , ⁹⁹ , ¹⁰⁰ , ¹⁰¹ , ¹⁰² , ¹⁰³ , ¹⁰⁴ , ¹⁰⁵ , ¹⁰⁶ , ¹⁰⁷ , ¹⁰⁸ , ¹⁰⁹ , ¹¹⁰ , ¹¹¹ , ¹¹² , ¹¹³ , ¹¹⁴ , ¹¹⁵ , ¹¹⁶ , ¹¹⁷ , ¹¹⁸ , ¹¹⁹ , ¹²⁰ , ¹²¹ , ¹²² , ¹²³ , ¹²⁴ , ¹²⁵ , ¹²⁶ , ¹²⁷ , ¹²⁸ , surged to represent 45% of studies in 2024 (n=56) ¹²⁹ , ¹³⁰ , ¹³¹ , ¹³² , ¹³³ , ¹³⁴ , ¹³⁵ , ¹³⁶ , ¹³⁷ , ¹³⁸ , ¹³⁹ , ¹⁴⁰ , ¹⁴¹ , ¹⁴² , ¹⁴³ , ¹⁴⁴ , ¹⁴⁵ , ¹⁴⁶ , ¹⁴⁷ , ¹⁴⁸ , ¹⁴⁹ , ¹⁵⁰ , ¹⁵¹ , ¹⁵² , ¹⁵³ , ¹⁵⁴ , ¹⁵⁵ , ¹⁵⁶ , ¹⁵⁷ , ¹⁵⁸ , ¹⁵⁹ , ¹⁶⁰ , ¹⁶¹ , ¹⁶² , ¹⁶³ , ¹⁶⁴ , ¹⁶⁵ , ¹⁶⁶ , ¹⁶⁷ , ¹⁶⁸ , ¹⁶⁹ , ¹⁷⁰ , ¹⁷¹ , ¹⁷² , ¹⁷³ , ¹⁷⁴ , ¹⁷⁵ , ¹⁷⁶ , ¹⁷⁷ , ¹⁷⁸ , ¹⁷⁹ , ¹⁸⁰ , ¹⁸¹ , ¹⁸² , ¹⁸³ , ¹⁸⁴ . This rapid rise made LLMs the most frequently studied architecture in 2024, surpassing rule‐based systems (41%), despite the absolute number of rule‐based studies remaining relatively consistent in 2023‐2024. This trend indicates a decisive shift towards investigating advanced generative models within the field.

Distribution by evaluation methodology and architecture

Research effort across the evaluation stages was primarily focused on clinical efficacy testing (n=75), followed by pilot feasibility testing (n=72), and foundational bench testing (n=13) (see Figure 3).

Figure 3.

Distribution of studies by evaluation methodology and chatbot architecture. Percentages indicate the proportion of chatbot architectures within each evaluation methodology category. Subtotals and percentages are rounded to the nearest whole number, which may result in apparent summation discrepancies. ML – machine learning, LLM – large language model.

Chatbot architecture distribution varied markedly across these evaluation stages. Foundational bench testing was dominated by LLM‐based systems, which accounted for over two‐thirds (77%) of studies at this stage, with smaller contributions from machine learning‐based (15%) and rule‐based (8%) systems. Rule‐based architectures predominated in later stages, accounting for over half of both pilot feasibility studies (58%) and clinical efficacy trials (65%). The proportion of LLM‐based studies decreased substantially in these stages, accounting for only 24% of pilot feasibility studies and 16% of clinical efficacy studies. Machine learning‐based systems remained a minority across all stages, ranging from 15% in foundational bench testing to 19% in clinical efficacy testing.

This stark contrast between stages indicates that, while LLMs are the primary focus of early technical validation, rule‐based systems remain the principal architecture undergoing human testing and clinical evaluation.

Target conditions, functional purpose, and outcome measures of chatbot studies

Analysis of the studies' target conditions, functional purpose, and outcome measures revealed distinct patterns in the application and evaluation of different chatbot architectures (see Figure 4).

Figure 4.

Distribution of chatbot studies by target condition, functional purpose, and outcome measure. Within each subcategory, the left bar indicates the percentage distribution of chatbot architectures used, and the right bar shows the percentage distribution of evaluation methodologies employed. Subtotals and percentages are rounded to the nearest whole number, which may result in apparent summation discrepancies. ML – machine learning, LLM – large language model.

Examining target conditions, research most frequently addressed general mental well‐being (n=51), depression (n=50), and anxiety (n=41). Rule‐based systems were the predominant architecture for studies targeting depression (58%) and anxiety (62%), compared to LLM (20‐25%) and machine learning‐based (13‐23%) systems. LLM‐based systems showed higher relative representation in studies targeting general mental well‐being (28%), compared with rule‐based (49%) and machine learning‐based (22%) approaches.

Methodologically, studies of general mental well‐being were largely in pilot feasibility testing (66%), with fewer in clinical efficacy testing (27%) or foundational bench testing (7%). By contrast, both depression and anxiety interventions had mostly advanced to clinical efficacy trials (57% and 58%, respectively), with smaller proportions in pilot feasibility (38% for depression; 33% for anxiety) and foundational testing (6‐8%).

Studies were grouped into five functional purposes, including emotional support (n=46), therapeutic interventions (n=42), education and skills training (n=27), assessment and monitoring (n=21), and general and other functions (n=24). In every category, rule‐based systems dominated, ranging from 48% in emotional support to 83% in general and other functions. LLM‐based approaches were most represented in emotional support (30%) and assessment and monitoring (29%), while machine learning‐based systems peaked at 27% for therapeutic interventions and dipped as low as 4% for general and other functions.

Evaluation methodology varied notably by functional purpose. Pilot feasibility testing was the most common method for assessment and monitoring (78%), emotional support (53%), and education and skills training (49%). By contrast, clinical efficacy testing led in therapeutic interventions (65%), and general and other functions (71%), with education and skills (43%) and emotional support (34%) also seeing substantial clinical work. Foundational bench testing remained minimal, accounting for 7‐13% of studies in four categories and 0% for general and other studies.

The choice of chatbot architecture and evaluation stage was associated with the outcome measures prioritized. Studies measuring clinical outcomes (n=99) predominantly employed rule‐based (59%) or machine learning‐based (18%) systems, with fewer using LLM‐based approaches (23%). These clinical outcome studies were most often evaluated via clinical efficacy testing (65%), followed by pilot feasibility (31%) and foundational bench testing (4%).

User experience evaluations (n=46) similarly favored rule‐based (64%) and machine learning‐based (19%) architectures over LLM‐based approaches (17%), and were overwhelmingly conducted as pilot feasibility studies (78%), with smaller proportions in clinical efficacy (17%) and foundational testing (5%). In contrast, technical performance studies (n=15) were dominated by LLM‐based systems (54%), with rule‐based and machine learning‐based approaches at 31% and 15% respectively, and were primarily assessed at the foundational bench (45%) and pilot feasibility (37%) stages, with only 18% reaching clinical efficacy testing.

Characteristics of evaluation stages

The nature of evaluations conducted on mental health chatbots evolved significantly across the research pipeline, particularly concerning the types of participants involved (see Figure 5).

Figure 5.

Evaluation participant types across study methodologies. Percentages indicate the distribution of participant types within each methodology. Subtotals and percentages are rounded to the nearest whole number, which may result in apparent summation discrepancies.

Foundational bench testing (n=13) primarily involved evaluations conducted with clinicians (62%) or represented technical assessments where participant type was “not applicable” (38%). Transitioning to the pilot feasibility testing (n=72), evaluation efforts focused predominantly on general users (78%) to assess usability and acceptability, alongside a substantial inclusion of patients (19%) for initial target population testing. The clinical efficacy testing (n=75) presented a more varied participant profile, characterized by the engagement of clinicians as evaluators (19%) and continued recruitment of general users (25%). Notably, explicitly identified patient participants were less frequently involved (5%) in this final stage compared to pilot studies, while a large proportion of these efficacy trials reported participant type as “not applicable” (37%) or “not specified” (13%), potentially reflecting diverse study designs or reporting practices.

During the pilot feasibility phase (n=72), study durations ranged from under one hour to two years (see Figure 6). Under one hour applied to eight studies, between one hour and one day to eighteen studies, and between one day and one week to sixteen studies. Ten studies did not report a duration. Rule‐based architectures appeared in half to two‐thirds of studies across every duration band. LLM‐based systems featured in 20% to 33% of studies, with a 31% share in the one‐day to one‐week group. Machine learning‐based approaches ranged from 0% in the not‐applicable category up to 25% in the one‐to‐four‐week interval. These results show that, although rule‐based chatbots dominate pilot feasibility testing, LLM and machine learning systems have also been trialed across the full spectrum of study lengths.

Figure 6.

Distribution of study durations for the various chatbot architectures in the 72 pilot feasibility studies. Percentages indicate the proportion of chatbot architectures within each study duration category. Subtotals and percentages are rounded to the nearest whole number, which may result in apparent summation discrepancies. ML – machine learning, LLM – large language model.

Terminology discrepancies: the meaning of “AI”

Beyond the core findings related to chatbot architectures and evaluation stages, the terminology used to describe these technologies in study titles also warranted examination (see Figure 7).

Figure 7.

Usage of “AI” in study titles versus actual chatbot architectures. The left bar shows the proportion of all studies containing “AI” in the title. The right bar shows the percentage distribution of underlying chatbot architectures for the subset of studies with “AI” in the title. AI – artificial intelligence, Rule – rule‐based, ML – machine learning, LLM – large language model.

Analysis revealed ambiguity in the use of the term “AI”. Of the 160 included studies, a small proportion (n=21, 13%) explicitly used the term “AI” in their titles; the vast majority (n=139, 87%) did not. Among the 21 studies that did use the “AI” label, the majority (57%) employed advanced LLM‐based systems. A further 19% utilized machine learning‐based approaches. Notably, however, the “AI” label was also applied in nearly one‐quarter (24%) of these cases to studies utilizing rule‐based architectures. As rule‐based systems operate on predefined scripts without the adaptive learning capabilities characteristic of contemporary machine learning and LLM systems, their inclusion under the general “AI” descriptor contributes to terminological ambiguity and potential misrepresentation of chatbot sophistication within the field.

DISCUSSION

The rapid adoption of generative LLMs (e.g., GPT‐4) reflects broader AI trends, but introduces unique risks in mental health contexts ¹⁸⁵ . The complexity of mental health conditions, the subjective nature of diagnosis, and the need for contextual understanding further complicate AI integration ¹⁸⁶ , ¹⁸⁷ , ¹⁸⁸ . While early rule‐based systems such as Woebot prioritized safety through scripted dialogues, LLMs such as Replika now risk generating unvalidated advice due to their reliance on uncurated datasets ¹⁸⁹ . This tension between innovation and safety, reflected in a December 2024 complaint by the American Psychological Association to the US Federal Trade Commission accusing a generative AI chatbot of harming children ¹⁹⁰ , underscores the need for structured validation frameworks and research to fill the gaps identified in our results.

A persistent challenge in the field is the misalignment between the marketed rhetoric of “AI‐driven” systems and their underlying technical realities. Platforms such as Woebot and Replika market themselves as “AI‐driven”, yet the term “AI” remains ambiguously defined and is often employed without clear specification of the underlying model architecture ¹⁹¹ .

Early iterations of “AI” chatbots predominantly operated through scripted, rule‐based interactions with only rudimentary machine learning enhancements. These rule‐based tools, emblematic of the good old‐fashioned artificial intelligence (GOFAI) paradigm, have been critiqued for their limited adaptability and depth, standing in sharp contrast to modern data‐driven, connectionist approaches ¹⁹² . As both technological capabilities and stakeholder expectations evolve, clinicians and patients now increasingly expect AI to represent more sophisticated, dynamic and autonomous connectionist systems, such as LLMs, capable of generating contextually rich, free‐form dialogue ¹⁹³ .

This evolving expectation creates confusion, as legacy systems continue to be marketed under the same broad AI umbrella, making it difficult to distinguish between LLM‐driven innovations and older rule‐based or hybrid approaches. To address this, our classification system offers a structured way to categorize mental health chatbots based on architecture and function, distinguishing rule‐based systems that operate through deterministic scripts, machine learning‐based systems that enhance adaptability through data‐driven models, and LLM‐based systems that generate free‐form, contextually rich dialogue. This architectural classification is critical, as it allows clinicians, researchers and regulators to appropriately evaluate chatbot capabilities, ensuring that expectations align with actual functionalities rather than misleading claims.

With the increasing number of chatbot studies and the sharp rise in LLM‐based chatbot research, the need for standardized evaluation frameworks is more urgent than ever. While chatbots were once predominantly rule‐based, requiring only basic assessments of functionality and engagement, the integration of machine learning‐ and LLM‐based systems has generated complexity in evaluation. Unlike rule‐based chatbots, which can be tested through predefined workflows, LLM chatbots introduce elements of unpredictability, requiring assessments beyond technical performance. Without a standardized nomenclature for evaluation, studies report outcomes inconsistently, making it difficult to compare findings across research. The introduction of a standardized three‐tier evaluation continuum – foundational bench testing, pilot feasibility testing, and clinical efficacy testing – addresses this need by providing a clear progression of evidentiary rigor. It also fits well with recent calls for graded regulation of LLM systems, with certification linked to the role of the chatbot and rigor of its real‐world efficacy ⁹ .

Architectural analysis reveals that the focus and stage of evaluation vary significantly depending on chatbot type. Rule‐based systems, lending themselves to structured interactions such as symptom monitoring or delivering psychoeducational content, continue to be the most common architecture evaluated in clinical efficacy trials (65% of T3 studies). This likely reflects their longer history and suitability for interventions where predictability and safety are paramount. Machine learning‐based chatbots, offering more adaptability than rule‐based systems but lacking the generative fluency of LLMs, are represented modestly across all evaluation tiers (15‐19%). In stark contrast, LLM‐based systems heavily dominate foundational bench testing (77% of T1 studies), indicating that current research primarily investigates their technical capabilities, such as conversational quality or adherence to specific prompts, often in simulated scenarios. Despite their potential for nuanced, high‐context interactions relevant to applications such as assessment or emotional support, LLMs are infrequently evaluated in T3 trials (16%). This suggests that, while LLMs are being actively explored for their technical promise, they have yet to undergo widespread, rigorous testing for clinical benefit in high‐stakes mental health contexts, leaving a critical gap in evidence.

A major challenge highlighted by the T1‐T3 framework is the heterogeneity in the types of evidence generated across the evaluation pipeline, often tied to the predominant chatbot architecture at each stage. Foundational bench testing (T1), where LLM‐based studies are most prevalent, typically yields evidence related to technical performance – such as conversational coherence, linguistic accuracy, or safety in controlled tests. As studies progress to T2 pilot feasibility testing, the focus shifts towards usability, engagement, and user acceptance, evaluated across diverse groups including general users and patients. Rule‐based studies are more common here (58%) than LLM‐based studies (24%). Evidence of clinically meaningful impact, such as symptom reduction measured by validated scales over time, is primarily generated in T3 clinical efficacy testing, and rule‐based systems are the main architecture assessed at this highest tier (65%).

The current concentration of LLM research in T1 and T2 stages means that these advanced models are often validated based on technical feasibility or short‐term user experience metrics, rather than demonstrated therapeutic effectiveness. This disparity underscores a crucial limitation: strong performance in T1 stage or positive user feedback in T2 stage does not necessarily translate to T3 clinical efficacy. Furthermore, while T2 studies often specify diverse participant groups, T3 evaluations show less consistency in reporting participant characteristics, sometimes hindering the assessment of real‐world applicability and long‐term impact.

Ethical, safety and regulatory concerns are becoming increasingly critical as chatbots move closer to clinical deployment. LLM‐based systems introduce considerable risks, including potentially greater data privacy violations than rule‐based or machine learning‐driven systems, algorithmic bias, and the potential for “hallucinations” (i.e., false or misleading responses) that could lead to harmful advice. All this may be exacerbated by the richer and more in‐depth conversational capabilities of LLMs, which could encourage users to disclose more sensitive information ¹⁹³ . Unlike rule‐based chatbots, which are constrained to predefined responses, LLMs rely on large, uncurated datasets, making them susceptible to misinformation. These risks are not hypothetical: real‐world examples, such as Replika's backlash for generating inappropriate responses ¹⁹⁴ , illustrate the consequences of insufficient safeguards in generative AI systems ¹⁹⁵ . While rule‐based chatbots mitigate some of these risks through structured outputs, they lack the adaptive empathy needed for sustained mental health support. Machine learning‐based systems occupy an intermediary position, balancing adaptability with limited generative capacity but often struggling with transparency and interpretability.

Addressing these risks requires regulatory bodies to establish clear certification pathways for LLM‐driven chatbots, ensuring that innovations in generative AI are balanced with accountability and user safety. However, given the widespread availability and increasing use of base models (e.g., GPT ⁸ , Gemini ¹⁹⁶ , Claude ¹⁹⁷ , Llama ²⁰ , DeepSeek ¹⁹⁸ ) for mental health applications, this alone may not be sufficient. An urgent research avenue is to independently establish the safety and clinical utility of these models, ideally through rapid, automated and repeatable evaluations as their capabilities continue to evolve.

Moving forward, mental health chatbot research must prioritize rigorous clinical efficacy trials for LLM‐based systems, ensuring that chatbots progress beyond foundational and feasibility testing to real‐world clinical validation. The development of standardized clinical endpoints, transparency in chatbot architectures, and regulatory alignment with AI‐driven mental health tools will be essential in bridging the gap between feasibility and efficacy. As chatbots continue to evolve, robust validation methodologies will be necessary to ensure that they serve as effective, ethical, and clinically reliable tools for global mental health care.

CONCLUSIONS

Mental health chatbots have rapidly evolved from deterministic rule‐based systems to sophisticated LLMs, signalling a transformative shift in digital psychiatry. Despite this promising advancement, our systematic analysis highlights a fragmented landscape with limited rigorous clinical validation, particularly concerning generative AI technologies. The proposed three‐tier classification system clarifies evaluation rigor and reveals that most LLM‐based interventions remain in early development phases.

Future research should prioritize rigorous clinical efficacy trials, transparent reporting of chatbot architecture, and ethical evaluations, to ensure that these technologies reliably enhance mental health care. Clinicians and policy makers must distinguish between marketing claims and technical realities, advocating for evidence‐based standards analogous to established medical AI certification processes.