This systematic review will answer research questions regarding the effects of AI-based interventions on anxiety and depression. The three research questions include:

Eligibility criteria were developed based on the PICOS (Population, Intervention, Outcome, Comparator, and Study Design) framework. ²⁸ The population in this study is the general population, such as students, parents, patients, adults, and workers, considering that depression and anxiety are psychological conditions that are cross-demographic and not limited to specific clinical groups, thus enabling the evaluation of the effects of AI interventions in various real-world contexts for various users. The intervention criteria focus on AI, including AI chatbots, ChatGPT, AI-based platforms, and other AI-based interventions, to distinguish AI’s effects from that of conventional interventions. The comparator in this study was non-AI interventions or alternative AI-based designs used as a control condition. Outcome measures focus on psychological problems, namely anxiety and depression, as these two conditions are the most common psychological disorders and are most often targeted in AI-based interventions in the literature. We filtered the article based only on the study using a Randomised Controlled Trial (RCT) to maximise internal validity. Criteria excluded from the study were research that focused solely on reducing depression or anxiety; interventions that were not AI-based, such as digital or conventional interventions; cross-sectional studies; experimental studies other than RCTs; quasi-studies; case reports; non-empirical articles; or articles written in languages other than English.

The literature search was conducted across four main databases: Scopus, CINAHL (EBSCO), Web of Science (WoS), and MEDLINE (PubMed), all of which are indexed in Scopus. The search was conducted over the last 10 years (6 November 2015–6 November 2025) to ensure the studies remain relevant today. The literature search focused on articles relevant to the PICOS framework that discussed the population (general), intervention (AI-based), outcomes (depression and anxiety), and study design (experimental, RCT). The Boolean operators used were (“artificial intelligence” OR AI) AND (depression) AND (anxiety) AND (“randomised controlled trial”). In addition, the article search was limited to English-language articles. The inclusion and exclusion criteria are described in Table 1.

Study selection was conducted by five reviewers (NMAPP, IGAAIRPD, RSA, KPNT, and MWNTN) using Rayyan AI tools. Articles obtained from four databases were then imported into Rayyan and automatically deduplicated. Initial screening was conducted by four reviewers (NMAPP, KPNT, IGAAIRPD, and MWNTN), filtering abstracts for suitability to the PICOS framework. This was followed by full-text screening, yielding 14 final articles. When disagreements arose, the senior reviewer made the final decision on which articles were included (RSA).

The extraction stage was carried out using Notebook LM and manual extraction by the author. Data extraction was carried out by identifying 14 selected articles and extracting publication information (author name, year of publication, country, study design, sample size, Scopus quartile), population characteristics, sampling techniques, interventions (duration and type), and findings. Data synthesis in this study was conducted using a narrative approach to interpret and integrate the findings, due to the conceptual and methodological diversity of the included studies. ²⁹ The synthesis involved organizing study characteristics and results into comparative tables, identifying similar and recurring patterns, and examining relationships among interventions across the literature. ²⁹

Data synthesis was performed narratively by categorising the findings into three research questions: the characteristics and patterns of research on AI interventions for anxiety and depression over the last 10 years; the effects of AI-based interventions; and the implementation of AI-based interventions for anxiety and depression. The extraction was presented in tables and descriptive narratives. Reliability assessment was carried out by ensuring the suitability of articles according to the inclusion-exclusion criteria, conducting systematic selection using PRISMA guidelines, extracting data using a standard format, and providing direct citations for all articles. The 14 selected papers are described in Table 2.

The risk of bias in the selected studies was assessed using the Cochrane Risk of Bias 2 (RoB 2), which is designed to assess randomised controlled trials (RCTs). The assessment was conducted on five main domains, namely: (D1) bias from the randomisation process, (D2) bias due to deviations from the planned intervention, (D3) bias due to missing outcome data, (D4) bias in outcome measurement, and (D5) bias in the selection of reported results. For each domain, each study was then categorised as low risk, some concerns, or high risk. ³⁰

The classification was determined based on the signalling questions in RoB 2, including the information available in the research report. Then, the overall risk-of-bias assessment was conducted in accordance with the official RoB 2 guidelines, ensuring that the final decision remained consistent and accountable. To maintain consistency in the assessment, the assessment process was carried out systematically by recording the reasons behind each decision in each domain (e.g., whether the randomisation procedure was described, whether there was potential for intervention deviation, and so on). If unclear information was found in an article, it was noted as a consideration in determining the relevant domain category, in accordance with the principle of caution in RoB 2. ³¹

Artificial intelligence (AI) holds potential for reducing symptoms of depression and anxiety due to its accessibility and complementary role in conventional care. Various AI tools, such as Large Language Model (LLM)-based agents, chatbots, and mobile applications have shown effects in reducing symptoms of depression and anxiety. ²⁶ Although AI shows significant effects in addressing mental health issues, it is not a substitute for professionals or therapists; rather, it is a complement, while safety and long-term effects still need to be considered. ²⁶ The effects of AI from the 14 studies included in this systematic review will be outlined in Table 4.

Most studies suggest that AI-based interventions may help reduce symptoms of depression and anxiety, although the evidence remains heterogeneous. ^{6, 18, 32– 36} Several controlled trials reported improvements in both outcomes; for example, ChatGPT-4.0 in digital counselling for cancer patients was associated with significant reductions in depression and anxiety compared with a control group. ³² However, findings were not consistent across all studies. Some interventions were effects for only one outcome, such as Psy-Bot for depression but not anxiety, ¹⁷ while ChatGPT in preoperative education reduced anxiety but not depression. ³⁷ In addition, several studies reported improvements only within groups, with no significant differences compared with control conditions. ^{21, 38} Overall, these findings indicate that AI-based interventions may be potential, but their effects appear to vary depending on population, intervention type, and study context.

Primary outcomes

The primary focus of this study was to measure the effects of AI interventions in reducing symptoms of mental disorders. Thus, the main outcomes were as follows: (1) reduction in depression symptoms measured using clinical scales such as the Patient Health Questionnaire-9 (PHQ-9), Centre for Epidemiologic Studies Depression Scale (CES-D), Hospital Anxiety and Depression Scale (HADS-depression), and Edinburgh Postnatal Depression Scale (EPDS), (2) reduction in anxiety symptoms measured using instruments such as the Generalised Anxiety Disorder-7 (GAD-7), HADS-Anxiety, State-Trait Anxiety Inventory (STAI), and perioperative anxiety, (3) clinically meaningful changes, namely by assessing whether AI-based testing tools can provide improvements in symptoms that are equal to or even greater than those achieved with conventional interventions.

Secondary outcomes

Secondary outcomes in this study extended beyond the use of AI in reducing symptoms of depression and anxiety, providing additional insights into improving life satisfaction and general well-being, as reported by 14.2% (2/14) of studies. A study conducted by Karkosz et al. ¹⁹ revealed that the “Fido” application was not only effective in relieving anxiety symptoms but also significantly helped participants feel more satisfied with their daily lives. Other reported outcomes in this study included reduced loneliness, improved mood regulation, and enhanced social functioning, primarily among students following a short-term chatbot intervention. ^{17, 19, 34, 35}

A systematic literature search was conducted ²⁷ using four sources, namely Web of Science, Scopus, PubMed, and EBSCO, yielding 355 articles. Deduplication was performed using Rayyan AI. 287 articles underwent title and abstract screening, of which 270 articles were excluded for not meeting the inclusion criteria, such as inappropriate study design (n = 59), irrelevant interventions (n = 124), not focusing on depression and anxiety (n = 39), retracted articles (n = 1), and review articles (n = 47). A total of 17 articles were read in full, but 3 articles with irrelevant study designs (n = 2) and high risk of bias (n = 1) did not meet the criteria. Thus, 14 studies were included in the final analysis, as illustrated in Figure 1. PRISMA Flow diagram of study selection.

This section presents the main characteristics of the studies included in the systematic review, providing an overview of the research context analysed. The presentation of these characteristics is an important component of systematic reviews, helping to understand variations in study design, population, and interventions that underlie the interpretation of results, ²⁷ as well as to explain the results of individual studies, as highlighted in previous systematic reviews. ³⁹ Table 3 summarises the research design, sample size, population characteristics, and sampling techniques, types of AI-based interventions, duration, and main outcomes reported.

The study designs of the 14 articles were predominantly randomised controlled trials (RCTs) (n = 14), encompassing variations such as two-centre RCTs, ³² single-centre RCTs, ²⁰ and two-arm RCTs. ^{19, 33} Some studies were designed as quasi-RCT or pilot RCT designs, ^{37, 38, 40} while other studies used RCTs. ^{6, 17, 18, 21, 34, 35}

Fourteen articles published between 2015 and 2025 consistently examined the effects of AI-based interventions in reducing depression and anxiety. The distribution of publications across years was as follows: 2018 (7%), 2021 (7%), 2022 (7%), 2023 (15%), 2024 (7%), and the majority in 2025 (57%). Four studies were conducted in the United States ^{6, 18, 20, 34} and one study was conducted in Argentina. ⁴⁰ Studies in Europe were conducted in Poland ¹⁹ and Italy. ²¹ Studies in Asia were conducted in Turkey, ³² Hong Kong, ³⁸ and China. ^{17, 35– 37} Oceania was represented by one study in Australia. ³³ Figure 2(A) presents the distribution of publication years based on 14 selected journals from 2015 to 2025. Meanwhile, Figure 2(B) illustrates the geographical distribution of studies on AI-based interventions for depression and anxiety between 2015 and 2025.

The sample sizes across the 14 studies ranged from small to moderate. Moderate-sized samples included more than 500 participants (n = 865), ³⁵ while other studies involved fewer than 500 participants. Gender distribution varied across studies, with most studies stated that females were the dominant population. The age range spanned from adolescents and young adults (university students) to adults and the elderly. The populations included were heterogeneous, such as students, patients with specific medical conditions (e.g., cancer, knee osteoarthritis, postpartum mothers), individuals with specific psychological problems (eating disorders, depression and anxiety, and work-related stress), and general populations such as parents, adults, and workers.

These digital interventions take various forms and are designed to address the limitations of traditional mental health services. The identified digital interventions include: (1) chatbots and conversational agents, which are the most common forms, including text-based applications such as AI Chatbot, Tess, Woebot, Psy-Bot, Fido, Therabot, and ED ESSI ^{17– 20, 33, 34, 36, 38, 40} ; (2) large language models (LLM), which are technologies such as ChatGPT (version 4.0) which are used as digital counselling agents or companions to provide medical information and emotional support ^{32, 35, 37} ; (3) integrated AI platforms, such as the Eleos Health system, which supports conventional therapy by monitoring patient progress and improving therapist efficiency ⁶ ; and ⁴ passive behaviour monitoring systems, which detect early signs of psychological stress through passive digital behaviour tracking. ²¹ The visualisation of interventions from the selected articles is presented in Figure 3.

The duration of AI-based tool use in the review was categorised into three time frames: (1) short term (7 days to 4 weeks), in which interventions were designed to provide rapid emotional support or triage; for example, Psy-Bot was used for 7 days, Tess for 3–4 weeks, and LLM-based chatbots for 28 days. Interventions using Socratic questioning and Therabot were also conducted for a duration of 2–4 weeks. (2) Medium term (6 weeks to 3 months), which is typically used to assess more stable clinical effects; for example, Woebot was used for 6 weeks, Tess and the TEO platform for 8 weeks, and the Eleos Health platform for 2 months. The use of ChatGPT 4.0 in a medical context (e.g., cancer and orthopaedic patients) was generally used for 3 months. (3) Long-term (more than 4 months) which involves more complex or monitoring-based with longer durations; for example, the Neil chatbot (16 weeks), and the ED ESSI chatbot (over 4 months). The duration of interventions from the 14 selected articles is visualised in Figure 4.

Figure 5(B) summarises the risk-of-bias assessment for the 14 trials included in this review. Overall, 9 studies were classified in the “some concerns” category (64.3%), 3 studies were assessed as low risk (21.4%), and 2 studies (14.3%) were judged to have a high risk of bias. These findings indicate that although the available evidence is generally potential, several studies still present methodological limitations that should be interpreted with caution.

Across domains, the most notable limitation was bias arising from deviations from intended interventions (D2), which was the only domain contributing to the high-risk ratings in this review. In contrast, bias due to missing outcome data (D3) was less problematic, with most studies classified as low risk in this domain. For the remaining domains—bias arising from the randomisation process (D1), bias in measurement of the outcome (D4), and bias in selection of the reported result (D5)—the most common judgement was “some concerns”, generally reflecting incomplete or unclear reporting of methodological procedures rather than clear evidence of serious bias. The detailed distribution of risk-of-bias judgements is presented in Figures 5(A) and 5(B).

Most studies employed passive control conditions (e.g., usual care or waitlist), which may limit the ability to control for placebo effects. Only a small number of studies used active control groups (e.g., psychoeducation, books, or nurse hotlines), ^{19, 38, 40} which may reduce inferential strength. Moreover, only one study involved a therapist in a face-to-face setting when delivering AI-based CBT, ⁶ while another study involved direct responses from a physician. ³⁷ Human support, however, appears to play an important role in influencing adherence and intervention effects.

Digital and social media–based recruitment methods tend to attract self-selected, technologically literate populations. As a result, there is a risk of selection bias, whereby the findings may not fully represent populations with lower digital literacy or those with limited access to devices due to economic constraints. In addition, study samples are often drawn from a single population segment. For example, postpartum studies may include only women without severe depression, while eating disorder studies may recruit only adolescents on waiting lists, thereby limiting generalisability. Furthermore, follow-up periods are relatively short, typically ranging from 2–8 weeks, making it difficult to assess long-term effects.

Heinz et al. ¹⁸ highlighted another issue related to engagement, characterised by a decline in user participation over time (i.e., low retention). This pattern is often attributed to a “novelty effect”, which diminishes after the initial sessions. In addition, Sharp et al. ³³ indicated challenges in integrating chatbots into standard clinical care systems, particularly for patients on waiting lists. Other findings suggest that the transition from rule-based chatbots to those based on large language models (LLMs), such as ChatGPT, introduces new challenges related to personalisation and safety. Although generative AI enables more natural interactions, concerns regarding data privacy and the potential for medical “hallucinations” remain prominent in implementation within formal healthcare settings. ³²

This systematic review has several clear strengths. First, its compilation follows PRISMA guidelines, making the review more transparent, organised, and easier to trace. Second, this review deliberately focuses only on RCTs, yielding higher-quality evidence than when study designs are mixed. Third, the review focuses not only on mental health in general, but specifically examines AI interventions that target two outcomes simultaneously: depression and anxiety. Fourth, the scope of AI interventions is also quite broad, ranging from chatbots and large language models to integrated AI platforms, machine learning-based prediction systems, and passive behaviour monitoring. This prevents the interpretation of results from being too “narrow” to a single technology type. Finally, this review does not stop at summarising the results; it also includes a formal assessment of study quality using the Cochrane Risk of Bias 2 (RoB 2) tool. By assessing potential bias across five domains, the reported results are more “fair” to read, making it easier to identify which findings are strong and which require more careful interpretation. This summary of strengths is then clarified through the visualisation in Figure 6.

Although this review has several strengths, it also has several limitations that researchers need to acknowledge. First, the included studies show considerable heterogeneity in the types of AI interventions, exposure durations, outcome measurement instruments, and participant characteristics. Second, although all studies used an RCT design, implementation quality was not always consistent. The risk of bias assessment results show that most trials remain in the ‘some concerns’ category, and a small number are even at high risk. Third, the scope for generalising the findings also appears to be limited. The majority of trials were conducted in middle- and high-income countries, with participants who tended to be younger, mostly female, and well-versed in digital literacy. Fourth, most studies had relatively short to medium follow-up periods, so the sustainability of the effects has not been fully addressed. Finally, this review included only English-language publications, so there may be a language bias, and some relevant studies in other languages may not have been accessible. The study’s limitations will be illustrated in Figure 7.

Several recommendations were made in the studies included in this review. First, future research should involve larger samples and more diverse populations. Second, the effects of conventional therapy should be compared with that of technology-based, face-to-face complementary therapies, such as integrating virtual reality, teletherapy, website-based therapy, and other AI interventions. Third, longer intervention durations are accompanied by follow-up sessions to assess the intervention’s long-term effects. Fourth, attention to participant safety and to effects testing procedures conducted in accordance with strict protocols. Fifth, providing interventions for higher-risk and more severe clinical disorders. Sixth, exploring the sustainable impact of the interventions provided. Finally, increasing human involvement with AI to enhance treatment impact, user satisfaction, and intervention usefulness. Recommendations from the 14 articles included in the study are explained in Figure 8.