Introduction

Conversational agents (CAs), or chatbots, are digital systems designed to engage users in interactive exchanges through text, voice, or visual interfaces []. CAs are designed to simulate human-to-human interactions, reimagined as human-machine dialogues []. In recent years, CAs have been increasingly adopted in mental health settings for their potential to enhance access and engagement, sustaining the use of digital interventions [,]. When used in the context of mental health, CAs aim to deliver support and even psychological interventions, blurring the line between digital convenience and therapeutic care [-]. There are 2 approaches that are used to design chatbots: rule-based and machine learning (ML)–based []. As artificial intelligence (AI) continues to evolve at an unprecedented pace [], mental health CAs (MHCAs) have become both a frontier of innovation and a matter of ethical concern.

While early reviews highlight promising outcomes from CA-based mental health interventions, the evidence remains mixed and context-dependent [,,]. Studies have shown that MHCAs are valuable for conducting private conversations [], aiding in learning [], improving users’ well-being [], preparing them for interactions with health care providers [], and boosting their self-efficacy [,]. Reflecting this growing interest, the MHCA ecosystem has expanded into a multibillion-dollar market, with widely adopted tools such as Wysa, Woebot, and Youper [,,]. These systems commonly draw on evidence-based therapeutic approaches, including cognitive behavioral therapy (CBT), mindfulness, positive psychology, and psychoeducation, to deliver scalable mental health support to users [].

However, recent instances of critical failures of CAs to properly and ethically support end users have ignited public scrutiny [,], raising questions on how to design, evaluate, and implement these tools in the mental health domain. On some occasions, users have reported that their interactions with CAs were distressing [] or approximated sexual harassment [] or that the CA appeared self-centered [] or irritating [], especially when the user felt misunderstood []. Concerns related to the precision, trustworthiness, and privacy of CAs have been raised as potential obstacles to user engagement and acceptance []. Additionally, a growing body of research has begun to unpack MHCA risks, from synthesized case studies of harm [,] to empirical studies on harms in CA interactions [-] and theoretical exploration of MHCAs’ intrinsic risks [,,]. Together, these accounts suggest that while CAs may offer scalable mental health support, they also introduce new forms of vulnerability, which demand thoughtful, safe design, and robust ethical oversight.

Several frameworks exist to guide the implementation and evaluation of health care technologies, including the Proctor model [], the Consolidated Framework for Implementation Research [], and the RE-AIM (Reach, Effectiveness, Adoption, Implementation, and Maintenance) framework []. While useful for general health care interventions, these models do not specifically address the unique challenges of designing conversational AI tools [] such as MHCAs. In human-computer interaction (HCI) and AI literature, more targeted guidelines for human-AI interaction have been proposed. For example, Amershi et al [] developed 18 design guidelines derived from literature and industry practice, refined through heuristic evaluation and expert review, providing general guidance for AI-infused products. Similarly, Yang and Aurisicchio [] conducted interviews to construct 10 guidelines for voice assistants, emphasizing competence, autonomy, and relatedness, and recommending features such as transparent system capabilities, socially appropriate conversation design, customization, and data control. In terms of adapting clinical concepts for therapeutic CAs, Moore et al [] evaluated MHCAs from the lens of basic, crucial prerequisites for therapeutic professionals’ conduct, while Song et al [] used therapeutic alignment to interpret MHCA users’ experiences. Despite these contributions, there remains a lack of guidance and consistency on equitable, inclusive, and trauma-informed design specifically for CAs in mental health contexts.

Trauma-informed care (TIC) is a strength-based framework for service delivery that emphasizes understanding and responding to the widespread, disempowering effects of trauma []. It involves acknowledging the impact of trauma and intentionally responding in ways that support safety and avoid retraumatization []. Many individuals experience traumatic events throughout their lives, regardless of their diagnoses or presenting conditions [], making trauma-informed approaches foundational in mental health care []. Trauma-informed approaches aim to maximize physical, psychological, and emotional safety in all health care interactions [] and not only those explicitly focused on trauma while also fostering opportunities for empowerment, control, and healing through safe, collaborative patient-clinician relationships []. Originally proposed by the Substance Abuse and Mental Health Services Administration (SAMHSA), the leading federal agency addressing mental health services in the United States, the TIC framework includes the following 6 key principles: safety; trustworthiness and transparency; peer support; collaboration and mutuality; empowerment, voice, and choice; and cultural, historical, and gender issues [].

While the TIC framework was initially developed to enhance therapeutic experiences and outcomes in individual psychotherapy and inform organizational policies, its application to technology design is increasingly recognized. In recent years, TIC concepts have been extended to domains such as telehealth and computing [-]. Trauma-informed telehealth research provides strategies for clinicians to promote safety, trust, and support during virtual visits [], while trauma-informed computing emphasizes a sustained commitment to designing digital systems that acknowledge trauma and its effects [,]. These approaches offer guidance for creating online environments that are trauma-sensitive and prioritize user safety, agency, and emotional well-being [-].

Given that digital technologies can inadvertently trigger or amplify trauma [], establishing design practices that minimize technology-facilitated harm and retraumatization is essential. As the TIC framework aims to enhance therapeutic experiences and outcomes across individual psychotherapy and organizational policy [], extending its application to technology design, particularly in the context of AI-based MHCAs, is both relevant and necessary []. Applying the TIC framework to CA design has the potential to improve their effectiveness as mental health interventions while addressing known risks, including user codependence [,], limited capacity to interpret complex emotional or nonverbal cues [], and potentially harmful responses to sensitive disclosures [,,]. Consistent with these concerns, systematic reviews of mental health CAs have identified user safety [,,] and trust within the user-CA relationship [,,] as critical and ongoing priorities for current and future research.

While trauma-informed ideas have been discussed across various areas of computing, to date, there has been no systematic effort to apply them to MHCAs. This gap is noteworthy, given the high prevalence of trauma among individuals seeking support through digital mental health tools []. The absence of a trauma-specific evaluative framework limits our ability to assess whether current MHCA designs adequately promote safety; trustworthiness and transparency; collaboration and mutuality; empowerment, voice, and choice; sensitivity to cultural, historic, and gender issues; and peer support to protect end users with trauma histories. Although previous literature reviews have extensively examined the efficacy, usability, and safety of MHCAs [,,,,-], it remains unclear how, or to what extent, existing interventions align with or operationalize TIC principles.

We selected SAMHSA’s TIC framework as the guiding lens for this review, as it provides a robust and translatable [-] foundation for evaluating trauma-informed design in digital contexts. By applying this framework, this review aimed to bridge a disciplinary gap between clinical care and technology design to provide a structured, trauma-aware evaluation of MHCA research and design to date. Accordingly, this scoping review maps how TIC principles are reflected, explicitly or implicitly, within existing AI-based MHCA research and identifies areas where trauma-informed approaches remain underused but could improve user experience and clinical outcomes. Our guiding research questions are as follows:

Which TIC principles are most frequently explored or integrated in the evaluation of MHCAs, and how are they operationalized?What key design considerations and recommendations are proposed in the literature for integrating TIC principles into MHCAs?Are there significant gaps in the literature regarding the application of TIC principles in CA technologies for mental health? If so, what areas require further exploration?
MethodsEligibility Criteria

We followed the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist ( to ensure the reliability of our results []. To meet the eligibility criteria, the studies had to (1) evaluate an MHCA for efficacy (either in symptom management or user experience), (2) center participants who were the main users of the MHCA, (3) only look at MHCAs that were independent mobile apps or web-based apps with no human involvement, (4) be published in peer-reviewed journals or conference proceedings in English between 2000 and 2024, and, finally, (5) include explicit or implicit references to principles from SAMHSA’s TIC framework [,,]. We did not register a review protocol prior to conducting the study.

We focused on mobile- and web-based MHCAs because these platforms represent the most accessible and widely used forms of digital mental health interventions in everyday contexts. This narrower scope aligns with prior reviews examining CAs for mental health [,], allowing for a more nuanced understanding of how trauma-informed design choices are implemented in the tools with which people most frequently engage. We did not include “social robots” or “embodied agents” as search terms but did exclude papers about physical robots, as their design architectures, usage contexts, and modalities differ from app- and web-based chatbots. Some papers used the term “embodiment” to refer to visual depictions or avatars of text- or audio-based CAs; we included these papers.

Articles were excluded if the MHCA was embedded in a preexisting communication platform (eg, Facebook Messenger), ensuring that the system was purpose-built for mental health intervention rather than serving as an ancillary feature. Extended abstracts and posters were excluded because they typically present preliminary or early-stage work that lacks the methodological and analytical depth needed to assess TIC principles. In contrast, we included both commercially available and research prototypes, provided they presented complete or well-documented evaluations relevant to TIC principles. This approach ensured that the review captured both mature systems currently in use and innovative prototypes that may inform future trauma-informed design in mental health technologies ().

Textbox 1. Eligibility criteria for scoping review.

Inclusion criteria:

Study focus: evaluation of mental health conversational agent (MHCA) for efficacy in symptom management or user experienceTrauma-informed care (TIC): mentions at least one principle from TIC frameworkUser interaction: primary users using MHCA for a mental health–related concernConversational agent (CA) type: independent mobile or web appCA design: has no human involvementStudy type: randomized controlled trial, quasi-experimental trials, experimental designs, user studies, pilot studies, observational research, and between-subject studiesArticle type: peer-reviewed articles, journals, and conference proceedingsLanguage: EnglishYear: studies published between 2000 and 2024

Exclusion criteria:

Study focus: not evaluating an MHCATIC: does not mention at least one principle from TIC frameworkUser interaction: CA is not meant for mental health or primary users are not using it for a mental health–related concernCA type: not an independent mobile or web app; a physical robotCA design: has human involvementStudy type: Wizard of Oz studies, systematic or scoping reviews, analyses of user reviewsArticle type: abstracts, extended abstracts, dissertations, editorials, position statementsLanguage: other languages than EnglishYear: studies not published between 2000 and 2024

To identify relevant articles, we searched Google Scholar, the Association for Computing Machinery (ACM) Digital Library, and PubMed in August 2024 for publications from 2000 to 2024. Our search terms for all 3 databases were as follows: (“mental disorder*” OR “mental health” OR “mood disorder*” OR autism OR “depression” OR “anxiety” OR phobia OR bipolar OR schizophrenia OR affective disorder OR psychosis OR psychotic disorder OR obsessive compulsive disorder OR panic disorder OR post-traumatic stress disorder OR substance abuse OR eating disorder) AND (“conversational agent*” OR “artificial intelligence*” OR “conversational AI*” OR “conversational bot*” OR “CAI*” OR “conversational system*” OR “conversational interface*” OR “smart-bot *” OR “virtual agent *” OR “virtual coach *” OR “avatar*” or “chatbot*” OR “chat bot*” OR “chatterbot*"). These search terms were based on previous studies on MHCAs [,-]. Although suicidality is a critical aspect of mental health, we did not include “suicide” or “suicidality” as explicit search terms. This decision followed prior scoping reviews on MHCAs [,], which focused on common diagnostic conditions, such as depression, anxiety, and post-traumatic stress disorder, domains where suicidality frequently co-occurs.

Following the initial retrieval of articles from the databases (N=25,857; ACM Digital Library: n=17,084, 66.07%; Google Scholar: n=220, 0.85%; and PubMed: n=8553, 33.08%), 4 authors (FFN, RP, ER, and KV) independently and manually screened the titles and abstracts to assess eligibility based on predefined criteria. The 4 authors divided the databases year-wise (ie, each was assigned a 5-y span). After this initial screening based on titles and abstracts, out of 25,857, a total of 99 (0.4%) articles were retained. Next, all 5 authors conducted relevancy coding by reviewing the full texts of these 99 articles based on the inclusion criteria. Relevancy coding was documented in Microsoft Excel. Discrepancies were discussed collectively and resolved through consensus. Of the 99 full-text articles, 81 (81.82%) were excluded—1 due to duplication and 80 for not meeting the inclusion criteria (user interaction: 37/80, 46.25%; article type: 5/80, 6.25%; CA type: 20/80, 25%; CA design: 2/80, 2.5%; study type: n=13/80, 16.25%; and no TIC elements: 3/80, 3.75%), resulting in 18 (18/99, 18.18%) studies. To broaden the scope of the review, the authors searched through these 18 papers’ citations to identify additional papers. A total of 62 potentially relevant references were manually screened. By consensus, 20 additional studies were included, resulting in a final sample of 38 papers for review. During screening, we used a broad, inclusive definition of references to the TIC framework that allowed us to explore how its core concepts were interpreted and operationalized across contexts without restricting the search to predefined terminology.

Data Extraction and Analysis

Five authors (FFN, FK, RP, ER, and KV) coded the data in Excel for a broad array of information on each publication in the corpus; publications were randomly assigned among authors. They extracted basic information per publication, including study methodology and outcomes and characteristics of the CA intervention, aligned with the PRISMA-ScR structure and previous studies []. Apart from the initial characterization of the studies, the analysis followed an abductive, theory-informed approach [], combining deductive sensitization to SAMHSA’s TIC principles with openness to inductively derived subthemes reflecting how these principles were operationalized across studies. As none of the papers explicitly referenced or applied SAMHSA’s TIC framework, identifying the 6 TIC principles required careful interpretive analysis. Descriptive and interpretive analyses of all extracted TIC-related data were conducted by the first and last authors, who iteratively revisited the data throughout the writing process to ensure that nuances in how TIC-related ideas were represented were accurately identified and not overlooked. To ensure methodological rigor and reduce subjectivity, the authors used a multistep, consensus-driven coding process. Discrepancies were systematically discussed and resolved through consensus, followed by iterative recoding to refine consistency and reliability. While this approach strengthened the credibility and reproducibility of our thematic interpretations, no qualitative work is neutral, and all interpretation was shaped by authors’ professional experiences and positionalities, as detailed in the next section.

Author Positionality

Embedded within a large health system, the interdisciplinary research team includes members with expertise in HCI (FK and FFN), social computing (FK and FFN), public and mental health domain (KV, RP, and FFN), health informatics (FFN, RP, and FK), and user experience research and design (ER, FFN, and FK). The team’s work is informed by ongoing collaboration with clinical providers and experience in engaging with mental health–related data, digital interventions, and sociotechnical systems in health care settings.

Clinical input from a licensed provider (a clinical psychologist from the same health system with expertise in trauma and TIC) was sought during the design of the study. Some team members have participated in trauma-informed technology design training within and outside of their health institution. However, our perspective is primarily grounded in applied informatics research and health care practice, shaped by close collaboration with clinicians and work in mental health settings. The team also includes researchers from both Western and non-Western backgrounds.

TIC Framework

The TIC framework, as defined by SAMHSA, provides a foundational approach for recognizing and responding to the impact of trauma across health care and other service systems []. Within this framework, trauma is defined as a combination of experiences that an individual perceives as harmful or life-threatening and the lasting adverse effects of those experiences on the individual’s functioning and well-being [,]. Rooted in research and expert consensus, the TIC framework is guided by six interconnected principles: (1) safety, ensuring physical and emotional safety in environments and interpersonal interactions; (2) trustworthiness and transparency, building trust through transparent, consistent, respectful, and fair communication and decision-making; (3) collaboration and mutuality, promoting partnership and reducing power imbalances between individuals, whether staff or clients; (4) empowerment, voice, and choice, recognizing and strengthening individuals’ existing capacities, voices, and experiences; (5) peer support, valuing and incorporating the perspectives and support of those with lived experiences of trauma; and (6) cultural, historical, and gender issues, being responsive to cultural, racial, historical, and gender-based contexts that shape individuals’ experiences [,]. These principles are designed to promote environments that acknowledge trauma’s pervasive effects, recognize its signs and symptoms, integrate this understanding into practice, and actively seek to prevent retraumatization []. We apply the TIC framework as a critical lens for evaluating mental health chatbots (MHCAs), as these digital tools often interact with users during moments of psychological vulnerability, and their design choices can either mitigate or exacerbate distress.

ResultsOverview of Included Publications’ Metadata (n=38)

represents the PRISMA-ScR checklist we followed. The study selection process is summarized in . A total of 18 (47.4%) publications were identified through the initial database search (PubMed: 9/38, 23.7% [,,-]; the ACM digital library: 7/38, 18.4% [-]; and Google Scholar: 2/38, 5.2% [,]; and ). The remaining 20 (52.6%) publications were found as citations of these 18 sources [-]. Overall, 30 out of 38 (78.9%) were journal articles, with the most frequent venues being the Journal of Medical Internet Research and affiliated journals (15/30, 50% articles), followed by Frontiers in Digital Health (4/30, 13.33% articles); 8 out of 38 (21.1%) were conference proceedings. Most (28/38, 73.7%) studies were published in 2020 or later, reflecting a recent surge in the domain. The United States was the most common study location (13/38, 34.2% studies), whereas 12 (31.6%) out of 38 articles involved users based in various European countries. While most studies’ participants were from a single country, 4 (10.5%) studies included participants from multiple countries [,,,], signaling some movement toward globally inclusive development.

‎Figure 1. PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) Flowchart outlining the identification and screening of publications for scoping review. TIC: trauma-informed care. Table 1. Basic characteristics of included publications (n=38).Parameters and characteristicsStudiesPublication metadata, n (%)Study designRandomized trial18 (47.4)Other experimental study types5 (13.2)User study141 (28.9)Pilot study67 (18.4)Survival analysis1 (2.6)Type of publicationJournal article30 (78.9)Conference proceeding8 (21.1)Publication sourceGoogle Scholar2 (5.3)PubMed9 (23.7)ACM Digital Library7 (18.4)Manual search20 (52.6)Study locationUnited States13 (34.2)China4 (10.5)Ireland3 (7.9)United Kingdom3 (7.9)Switzerland1 (2.6)Australia2 (5.3)Norway1 (2.6)Brazil1 (2.6)Argentina1 (2.6)Sweden2 (5.3)New Zealand1 (2.6)South Korea1 (2.6)Philippines1 (2.6)France1 (2.6)Finland1 (2.6)Scotland1 (2.6)Japan1 (2.6)The Netherlands1 (2.6)Belgium1 (2.6)Not available3 (7.9)Year of publicationPrior to 202010 (26.3)2020-202221 (55.3)2023-20247 (18.4)Sample characteristicsSample size, n (%)≤5011 (28.9)51-1006 (15.8)101-20012 (31.6)201-4994 (10.5)≥5005 (13.2)Age (y)Mean (range)32.52 (17‐69.2)Sex (% male)Mean (range)34.92 (0‐82)Recruitment setting, n (%)Clinical7 (18.4)Nonclinical29 (76.3)Clinical and nonclinical2 (5.3)

aPercentages were rounded and may not sum to 100.

bNumbers do not add up as some studies used more than one methodology.

cACM: the Association for Computing Machinery.

dNumbers do not add up as some studies took place in 2 or more countries.

eMean age was reported in 25 studies.

fSex percentages were reported in 32 studies.

A total of 18 (47.4%) studies included in this scoping review were randomized trials (randomized controlled trials: n=17, 94.4%), with 4 (22.2%) identified as preliminary or pilot studies. Sample sizes for randomized controlled trials ranged from 30 to 700 (mean 140.35, SD 153.99; median 107, IQR 58-148). Additionally, 5 (13.2%) studies used other experimental designs, including between-subject studies [,], single-arm pre-post intervention studies [,], and nonrandomized prospective studies []. Eleven (28.9%) publications were based on user studies and included mixed methods (n=6, 54.5%) [,,,,,], qualitative approaches (n=3, 27.3%) [,,], and quantitative approaches (n=2, 18.2%) [,]. Observational analyses of commercial app data (5/38, 13.2%) [,,,,] featured the largest sample sizes of all study types (mean 2334.8, SD 1775.8; median 2194, IQR 667-4073). Finally, 1 (2.6%) study comprised a quantitative survival analysis using data originally collected from a separate clinical trial [].

Overall, participant samples in the reviewed studies skewed younger (mean age 32.52, SD 14.6 y) and female (male participants: mean 34.92%, SD 18.64%). Twenty-nine (76.3%) publications involved only nonclinical samples, whereas 7 (18.4%) included clinical samples. Two (5.2%) studies recruited both clinical and nonclinical samples for comparison [,]. Depression and anxiety were the most frequently addressed concerns across the studies included [,,,,,,,,,,]. Summaries of publications can be found in ; more information is available in .

Overview of MHCA Interventions Reported in the Included Studies (n=28)

Most MHCAs were reported in a single publication, although a few appeared in multiple studies (eg, Woebot was included in 6 studies, and Wysa was included in 5 studies). While most publications evaluated a single MHCA, some did multiple MHCAs [,,]. Thus, 28 distinct MHCAs were represented across the corpus. Seventeen (60.7%) of 28 MHCAs were described as prototypes at the time of the study, and 9 (32.1%) were commercially available. However, close to half (17/38, 44.7%) of the publications in the corpus focused on commercial MHCAs. The most popular MHCAs were Woebot [,,,,,] and Wysa [,,,,]. Text (including emojis) or multiple-choice options were the most common input and output modalities, with 13 (13/28, 46.4%) MHCAs including other modalities such as audio or video. Sixteen (16/28, 57.1%) MHCAs used rule-based functionality for conversation logic, 1 (1/28, 3.6%) was fully generative AI, and 9 (9/28, 32.1%) incorporated both. Eighteen (18/28, 64.3%) MHCAs were described to have a visual avatar (eg, realistic animated face [] or nonhuman abstract character []). Fourteen (14/28, 50%) MHCAs were described as having a form of crisis intervention. While other MHCAs could have this feature, it was not directly mentioned in the articles.

One component of design frequently absent from included articles was the extent to which the MHCA versus the user guided the interaction, as well as examples of representative interactions. Out of 28 MHCAs included in this study, 26 (92.9%) delivered interventions to improve mental health; one (3.6%) focused on diagnosis [], and another (3.6%) focused on hospital discharge counseling []. Mental health improvement interventions were most often based on CBT (14/26, 53.8%), psychoeducation (8/26, 30.8%), mindfulness (4/26, 15.4%), positive psychology (3/26, 11.5%), motivational interviewing (2/26, 7.7%), self-care or self-help (3/26, 11.5%), acceptance and commitment therapy (2/26, 7.7%), gratitude (2/26, 7.7%), and mood tracking (2/26, 7.7%). Eleven (42.3%) of 26 CAs that provided mental health interventions were multimodal, involving 2 or more different approaches. Summaries of MHCA interventions are presented in , with more information on MHCAs provided in .

Table 2. Characteristics of mental health conversation agent (MHCA) intervention (n=28).Parameters and characteristicsChatbots, n (%)PurposeMental health intervention26 (92.8)Diagnosis1 (3.6)Hospital discharge counseling1 (3.6)StatusCommercial9 (32.1)Prototype17 (60.7)Not available2 (7.1)Response generationRule based16 (57.1)Artificial intelligence1 (3.6)Hybrid9 (32.1)Unclear from paper text2 (5.3)Input and output modalityText was sole input and output modality3 (10.7)Text and multiple choice were sole input options13 (46.4)Multiple choice was sole input option5 (17.9)Included emojis as input or output4 (14.3)Include audio or voice as input or output8 (28.6)Included infographics or images as input or output7 (25.0)Included video as input or output3 (10.7)Targeted disorderGeneral mental well-being3 (10.7)Depression and/or anxiety10 (35.7)Substance abuse disorders1 (3.6)Emotional distress or stress6 (21.4)Eating disorders2 (7.1)Schizophrenia1 (3.6)Chronic pain1 (3.6)Panic disorder1 (3.6)Posttraumatic stress disorder1 (3.6)Flexible or user-determined targeting2 (7.1)No specific targeted disorder7 (25)

aNumbers do not add up as several MHCAs had more than one input or output modality.

bNumbers do not add up as several MHCAs target more than one health condition.

Exploration of TIC Principles in Included PublicationsOverview

Although no publications cited the TIC framework, we explored both explicit and implicit references to its principles [,,]. This approach allowed us to capture trauma-informed practices that may be present but not formally acknowledged in the design and evaluation of the MHCAs, providing a more holistic picture. This echoes the scoping review by Eggleston et al [], which also found that while digital interventions neither used the term “trauma informed” in their described design processes nor cited SAMHSA, the authors could extract analyzable allusions to TIC principles. It is also important to note that SAMHSA is an American organization. Studies conducted outside the United States may be less likely to align with this framework, as trauma-informed practices can vary across international contexts.

We classified explicit references as instances where papers focused on one or more TIC principles using the exact principle names (eg, safety, trust and trustworthiness, and transparency) (). Implicit references to TIC principles reflected discussions of concepts that aligned with or alluded to TIC principles but were not specifically named (). As the purpose of this review was to evaluate the presence of TIC-aligned ideas across diverse literature, identifying implicit references required careful interpretation. To justify these determinations and limit subjective bias, we cross-checked each potential implicit reference against SAMHSA source texts [,,,], considered how the concept functioned within the paper’s stated aims and objectives, and engaged in iterative team discussions to reach consensus. The sections that follow outline how these explicit and implicit references were expressed, implemented, and measured.

Table 3. Explicit references to trauma-informed care (TIC) principles in included publications, including the name of the principle, the mental health conversational agent (MHCA) and publications that included that principle, and how the TIC-related consideration was measured or included.TIC principle and subprincipleMHCAImplementationSafety (n=8)SafetyEmohaa []Intervention designSafetyWoebot []Intervention designSafety3MR_2 []Intervention design and introductionSafetyCarebot []DiscussionPerceived safetyKIT []Finding from qualitative user studyPerceived safetyBotstar []Godspeed-V, introduction, and discussionPerceived safetyWysa []Discussion and intervention designPerceived safetyChatPal []Findings of supplementary qualitative user study and discussionTrustworthiness and transparency (n=18)TrustWysa []Definition of therapeutic alliance, associated with WAI-SR scaleTrustLaura []Scale response satisfaction questionnaire, introduction, intervention design, and findingTrustElizabeth []Definition of therapeutic alliance, associated with WAI-SR scaleTrustChatPal []Scale response questionnaireTrustWoebot []Finding from qualitative user study, introduction and discussionTrustEMMA []IntroductionTrustChatPal []Finding from qualitative user study and discussionTrustPhilobot []DiscussionTrustCarebot []Trust in Automation scale, findings of supplementary qualitative user study, and introductionTrustWoebot []Findings of supplementary qualitative user study and discussionTrustUnnamed []Associated with Acceptability E-Scale and discussionTrustChatPal []Findings of supplementary qualitative user studyTrustUser-chosen name []DiscussionTrustBotstar []Multi-Dimensional Measure of Trust, introduction, and discussionTrustXiaoE []Item in Working Alliance QuestionnaireTrustBella []Items in Friendship QuestionnaireTrust and transparencyChatPal []DiscussionTransparencyCarebot []Item in scale response trust questionnaire and discussionTransparencyWoebot []Finding from qualitative user studyTransparencySELMA []Intervention designTransparencyChatPal []Finding of supplementary qualitative analysis of user studyTransparencyUser-chosen name []Finding and discussionCollaboration and mutuality (n=11)CollaborationWysa []Item in WAI-SR and definition of therapeutic allianceCollaborationElizabeth []Item in WAI-SRCollaborationSELMA []Item in WAI-SRCollaborationXiaoNan []Item in WAI-SRCollaborationWoebot-SUDS []Item in WAI-SRCollaborationUser-chosen name []Item in WAI-SRCollaborationWoebot-SUDs []Item in WAI-SR and definition of therapeutic allianceCollaborationXiaoE []IntroductionCollaborationWysa []IntroductionCollaborationMultiple names []IntroductionCollaborationWoebot []DiscussionPeer support (1)Peer supportWoebot []Introduction and findings of qualitative user studyEmpowerment, voice, and choice (n=3)EmpowermentKIT []Finding from qualitative user studyEmpowermentSELMA []Intervention designEmpowermentChatPal []Finding from qualitative user studySensitivity to cultural, historical, and gender issues (n=2)Sensitivity to gender issuesKIT []Finding from qualitative user studySensitivity to cultural issuesChatPal []Introduction

aWAI-SR: Working Alliance Inventory–Short Revised.

Table 4. Summary of implicitly trauma-informed care (TIC)–related concepts in the corpus, including the concept name, how many publications included it, and where these concepts were found.Related TIC principle and implicit TIC conceptLocation of implicit TIC concept (n publications)Qualitative findings, n (%)Quantitative findings, n (%)System design or method, n (%)Discussion point, n (%)Other, n (%)Safety (n=32)Mental health crisis–related content option (n=4)0 (0)0 (0)3 (75)1 (25)0 (0)Crisis or MHCA failure detection and real-life services or hotlines provision (n=10)0 (0)0 (0)9 (90)1 (10)0 (0)Digital safety–related concepts (anonymity, integrity, password protection, security, and privacy; n=13)2 (15.4)1 (7.7)4 (30.8)4 (30.8)5 (38.5)Input handling for safety (limiting input options and ability to handle unexpected input; n=7)2 (28.6)0 (0)3 (42.9)2 (28.6)0 (0)Enable self-disclosure (n=12)5 (41.7)2 (16.7)0 (0)4 (33.3)4 (33.3)Positive emotional and psychological experience of use (eg, empathy, validation, nonjudgmental, and warmth; n=26)11 (42.3)7 (26.9)9 (34.6)10 (38.5)1 (3.8)24/7 availability (n=4)2 (50)1 (25)0 (0)1 (25)1 (25)Trustworthiness and transparency (n=29)