Trauma-informierte Chatbots in der psychischen Gesundheit: Nutzerperspektiven und Erfahrungen

TYPE Original Research PUBLISHED 10 June 2026 DOI 10.3389/fdgth.2026.1797681 Trauma-informed conversational agents for mental health: EDITED BY Michael Patrick Schaub, understanding user perspectives University of Zurich, Switzerland REVIEWED BY and experiences Warren Mansell, The University of Manchester, United Kingdom Xintong Lu1*, Shion Guha1,2,3, Rachel Pfafman1, Jessica A. Pater1, Marissa De Vries, Arkin, Netherlands Courtney L. Washington4,5 and Fayika Farhat Nova1 1 *CORRESPONDENCE Parkview Health, Parkview Research Center, Fort Wayne, IN, United States, 2Department of Computer Xintong Lu Science, University of Toronto, Toronto, ON, Canada, 3Faculty of Information, University of Toronto, xintong.lu@parkview.com Toronto, ON, Canada, 4Parkview Health, Parkview Behavioral Health Institute, Fort Wayne, IN, United States, 5Psychological Science Department, Northern Michigan University, Marquette, MI, United States RECEIVED 27 January 2026 REVISED 09 April 2026 ACCEPTED 16 April 2026 PUBLISHED 10 June 2026 Introduction: Mental health conversational agents (MHCAs) offer scalable, CITATION accessible psychological support yet raise concerns about safety and Lu X, Guha S, Pfafman R, Pater JA, appropriateness for trauma, exposed users. While trauma-informed care (TIC) Washington CL and Nova FF (2026) Trauma-informed conversational agents principles, emphasizing safety, trust, empowerment, collaboration, peer for mental health: understanding user support, and cultural sensitivity, are well-established in clinical practice, their perspectives and experiences. application and user interpretation in chatbot contexts remain unexplored. Front. Digit. Health 8:1797681. doi: 10.3389/fdgth.2026.1797681 This study aimed to explore (1) how users conceptualize trauma-informed care in the context of mental health chatbots and (2) what factors can help us COPYRIGHT © 2026 Lu, Guha, Pfafman, Pater, predict whether users perceive their chatbot interaction as trauma-informed. Washington and Nova. This is an open- Methods: A 59-item web-based survey (REDCap) assessed demographics, access article distributed under the technology proficiency, adverse life experiences, chatbot usage, and TIC terms of the Creative Commons Attribution License (CC BY). The use, perceptions among 606 participants recruited via ClickWorker, social media, distribution or reproduction in other and MyChart. Exploratory and confirmatory factor analyses identified latent forums is permitted, provided the TIC domains. Multivariable logistic regression determined predictors of original author(s) and the copyright owner(s) are credited and that the trauma-informed perception. original publication in this journal is Results: High satisfaction (84.5%) and trauma-informed perception (92.9%) were cited, in accordance with accepted reported. Factor analyses validated a five-domain structure: Trust & academic practice. No use, distribution or reproduction is permitted which does Transparency, Data Safety, Empowerment, Peer Support, and Cultural not comply with these terms. Sensitivity (CFI = 0.978; RMSEA = 0.045). Trust (OR = 3.89, p = .001), Empowerment (OR = 1.97, p = .025), and Peer Support (OR = 1.73, p = .021) significantly predicted trauma-informed perception. Convenience-driven use (OR = 0.33, p = .041) and smartphone proficiency (OR = 0.38, p = .033) showed negative associations. Discussion: This study provides empirical foundation for measuring trauma- informed design in MHCAs. Trust-building, emotional validation, and peer connection significantly shape user perceptions. Findings emphasize the necessity of aligning AI-based mental health tools with trauma-informed principles to enhance therapeutic safety, credibility, and inclusivity across diverse user populations. KEYWORDS artificial intelligence, chatbots, conversational agents, human-AI interaction, mental health, trauma-informed care Frontiers in Digital Health 01 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 1 Introduction (3) peer support; (4) collaboration and mutuality; (5) empowerment, voice, and choice; and (6) attention to cultural, Conversational agents (CAs), or chatbots, are AI-driven historical, and gender issues [27]. Given that many individuals systems designed to simulate natural human conversation experience traumatic events throughout their lives, regardless of through text or voice interfaces [1]. In mental health contexts, their diagnoses or presenting conditions, TIC has become a these technologies are increasingly used to deliver foundational approach in mental health care and has been psychoeducation, self-guided therapy modules, and supportive extended beyond traditional clinical settings to domains such as dialogue that mimics elements of traditional psychotherapy [2, trauma-informed telehealth and trauma-informed computing 3]. Several studies suggest that mental health conversational [28–34]. Critically, recent scholarship has raised concern that agents (MHCAs) can improve emotional well-being, reduce generative AI systems may actively increase the risk of symptoms of depression and anxiety, and enhance users’ self- technology-mediated trauma among vulnerable populations, efficacy and motivation to seek care [4–7]. These tools typically underscoring the urgency of applying trauma-informed draw on established therapeutic models such as cognitive principles to AI-enabled health tools [28]. A recent scoping behavioral therapy (CBT), mindfulness, and positive psychology, review of 118 studies systematically mapping TIC principles in and provide continuous access, anonymity, and convenience, digital technology design found that while safety, which may be particularly attractive for individuals with unmet trustworthiness, and transparency were the most commonly mental health needs [8–10]. As a result, commercial and referenced principles, a majority of studies invoked trauma- research-based MHCAs such as Woebot, Wysa, and Youper informed language without adherence to standardized have proliferated across digital marketplaces and health systems frameworks, revealing a persistent gap between TIC aspiration [8, 9, 11, 12]. and implementation in digital contexts [24]. These trauma- Despite their growing popularity, MHCAs raise questions about informed approaches aim to maximize physical, psychological, safety, ethical design, and therapeutic efficacy, particularly for and emotional safety in all health care interactions, fostering vulnerable users [13–16]. Reviews have noted that the evidence opportunities for empowerment, control, and healing through base remains mixed, with heterogeneous outcomes and high risk safe, collaborative patient-clinician relationships [31, 34]. of bias in many studies [4, 10]. Mental health chatbots span a Recently, scholars in human-computer interaction and digital heterogeneous landscape of systems, from earlier rule-based and health have advocated integrating TIC into the design of digital scripted conversational agents to contemporary large language platforms and AI-enabled technologies, giving rise to the model (LLM)-based generative AI systems such as those concept of trauma-informed computing [29, 35–37]. Parallel underpinning Woebot and Wysa [17, 18]. While earlier systems developments in companion robot design have similarly relied on predefined scripts and limited natural language demonstrated the feasibility of embedding TIC principles, understanding, generative AI systems introduce distinct including safety, trust, self-compassion, and self-efficacy, into AI challenges, including the potential for plausible but clinically system design for trauma-affected populations, providing an inappropriate responses, inconsistent emotional attunement, and important proof of concept for trauma-informed AI limited accountability [19, 20]. Empirical comparisons of LLM- development more broadly [25].A trauma-informed chatbot generated responses with those of licensed therapists have found would not only protect user privacy and emotional safety, but that chatbots tend toward directive, advice-giving communication also communicate its capabilities and limitations clearly, support rather than the Socratic, exploratory dialogue characteristic of user autonomy over how and when to engage, and avoid effective psychotherapy, and are currently considered unsuitable language or interaction patterns that may be experienced as as standalone therapeutic agents [20]. A scoping review of 101 coercive, dismissive, or triggering [29]. Despite growing studies further identified privacy and confidentiality (61% of advocacy for trauma-informed design, a critical gap remains: articles) and safety and harm (52%) as the most frequently existing research has primarily approached TIC from a design or discussed ethical concerns in conversational AI mental health developer perspective, with limited empirical attention to end- research [16].High-profile incidents, including reports of user experiences [24, 33, 38]. In particular, there is no well- inappropriate responses to suicidal ideation or encouragement of established, user-centered evidence base that systematically harmful behaviors, have heightened public concern and exposed examines how individuals perceive and respond to TIC critical limitations in design and oversight across both rule-based principles in MHCA interactions, nor which design elements and AI-generated conversational systems [21–23]. most strongly influence perceptions of psychological safety, Trauma-informed care (TIC) has emerged as a foundational trust, and empowerment in automated contexts. The present framework for addressing safety and ethical considerations in study directly addresses this gap by empirically investigating both clinical and digital mental health contexts. Originally end-user perceptions of trauma-informed care in MHCA proposed by the Substance Abuse and Mental Health Services interactions. Specifically, we examine: Administration (SAMHSA), TIC is a strengths-based approach to service delivery that emphasizes understanding and 1. RQ1: How do users conceptualize TIC principles in the responding to the widespread, disempowering effects of trauma context of mental health chatbots? [24–26]. Rather than functioning as a discrete treatment 2. RQ2: What factors predict whether participants perceive modality, TIC involves acknowledging the impact of trauma and their chatbot experience as trauma-informed? intentionally responding in ways that support safety and avoid retraumatization [27]. The framework is structured around six By centering user experiences and expectations, this study aims to core principles: (1) safety; (2) trustworthiness and transparency; clarify what trauma-informed care looks like in MHCA Frontiers in Digital Health 02 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 interactions and to identify actionable features that support open-ended questions. To ensure data quality, the survey perceived emotional safety, trust, and autonomy. The findings included one attention-check item and two hidden bot-detection provide insights for developers, clinicians, and researchers questions designed to identify inattentive or automated seeking to implement TIC in digital mental health tools and responses. While participation in all questions was not ensure that MHCAs serve users in an ethical, supportive, and mandatory and respondents could skip items at their discretion, trauma-sensitive manner. they were required to provide responses to demographic items and multiple-choice questions related to TIC. Participants 2 Materials and methods received $10 in compensation for survey completion. Full survey items are provided in Supplementary Table S1. 2.1 Survey design This study was conducted as part of a larger federally funded 2.2 Recruitment research project and was approved by the institutional review board (IRB) of the Parkview Health. A structured, self-administered Participants were recruited between December 2024 and April online questionnaire was developed and administered via REDCap 2025 through three primary channels: the crowdsourcing platform to assess users’ experiences with MHCAs, their perceptions of Clickworker (n ¼ 488), online social media platforms (n ¼ 113), trauma-informed chatbot characteristics, and their overall and the MyChart patient communication portal (n ¼ 5), satisfaction with chatbot interactions. yielding a final eligible sample of 606 individuals [43]. Given the lack of established instruments that operationalize Participation was voluntary, and each participant received $10 as trauma-informed care principles through end-user perspectives compensation for their time. specifically in the context of human–AI chatbot interactions, we Clickworker, a trusted crowdsourcing platform, was selected developed an instrument by mapping SAMHSA’s TIC as the primary recruitment channel given its access to a large, framework to measurable constructs. The items were developed diverse, and verified workforce spanning a broad range of through conceptual grounding in prior TIC and related HCI/ educational backgrounds [43]. Crowdsourcing platforms, such as healthcare literature, followed by iterative refinement based on Clickworker, are widely used in human-computer interaction expert feedback to improve clarity and relevance to chatbot and digital health research to recruit diverse, scalable participant experiences [27, 31, 39–41]. samples [44]. To address common data-quality concerns The questionnaire was developed as a self-designed instrument associated with crowdsourced samples, several safeguards were and refined with inputs from an expert in clinical psychology with implemented, including explicit attention-check items (e.g., “To expertise in TIC (the health system the authors are from), as well as make sure you are paying attention, please select the third feedback from human-centered design experts with experience in option for this question”), embedded bot-detection questions mental health contexts and AI ethics. These inputs helped (e.g., “What is your favorite color?” and “What is your favorite improve the clarity and relevance of the items while ensuring ice cream flavor”), and post hoc exclusion of respondents who alignment with trauma-informed principles. Likert-scale response reported using tools that did not meet the definition of a mental formats were used to allow participants to indicate the degree of health chatbot. Only participants who successfully passed all their experiences or agreement in a consistent and interpretable quality checks were included in the final analytic sample. manner across constructs. Item wording was adapted to reflect Social media recruitment was conducted via Twitter and user-facing experiences with chatbots rather than clinical or LinkedIn, where study information and a survey link were provider-oriented language. [26, 27, 34, 42]. publicly posted. The same data-quality safeguards applied to the The survey comprised four primary content sections: crowdsourced sample were implemented for social media respondents. 1. Demographics, which captured participants’ gender, age, Recruitment via MyChart targeted Indiana-based patients race/ethnicity, education level, and history of adverse life aged 18 years or older who had at least one clinical encounter experiences; within the behavioral health or psychiatry departments of the 2. Technology Experience, which assessed participants’ authors’ health system in the preceding 12 months and an active proficiency with smartphones, artificial intelligence, and behavioral or psychiatric episode documented in their record. smart devices, as well as their motivations for using Although recruitment for MyChart spanned approximately eight mental health chatbots; weeks, the resulting subsample (n ¼ 5 participants) was small 3. Trauma-Informed Chatbot Characteristics and and is therefore not interpreted as representative of a clinical Experiences, which evaluated user perceptions of TIC population, likely reflecting the sensitivity of trauma-related principles as reflected in chatbot design and interaction research topics and the well-documented reluctance of clinical quality; and populations to participate in such studies 4. Satisfaction Level, which measured overall satisfaction The study was not designed to exclusively recruit trauma- and perceived trauma-informed quality of the chatbot affected or help-seeking clinical populations; instead, interaction. recruitment was intentionally broad to reflect the real-world diversity of individuals, both with and without trauma exposure, The survey in total includes 59 items, including multiple-response who engage with digital mental health tools. This approach (checkbox), Likert-scale (rated from 1 = not good/proficient/never aligns with an increasingly recognized perspective in human- to 5 = very good/proficient/always), binary (yes/no), and computer interaction and mental health technology research Frontiers in Digital Health 03 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 that such heterogeneous user populations are a valid and 2.4.1 Descriptive analysis important focus of study. Accordingly, the study is framed as exploratory and user-centered rather than intended for Descriptive statistics were computed to summarize participants’ clinical generalization. overall satisfaction, technology proficiency, chatbot use and motivations, as well as multiple-response items on TIC perceptions. Frequency distributions and proportions were reported for categorical and binary variables, while means and standard 2.3 Participants deviations were calculated for continuous and ordinal variables. A total of 606 individuals constituted the final eligible analytic 2.4.2 Factor analysis sample. Eligibility required participants to be: (1) aged 18 or older; (2) self-reporting a history of mental health concerns; (3) having To identify the underlying structure of TIC perceptions, an used at least one mental health chatbot; and (4) able to read and EFA was conducted on the full set of Likert-scale items from the understand English. The informed consent item was displayed TIC section of the survey. EFA was chosen given the exploratory only to individuals satisfying all four criteria. Those who failed aim of identifying latent constructs underlying participant eligibility screening, quality checks, or provided invalid chatbot responses, as no established factor structure was assumed. Prior responses were excluded. to analysis, all variables were assessed for missing data. The Participants were instructed to reflect on their most recent or average item-level missing rate was , 2%. To preserve the most frequently used mental health chatbot, acknowledging that integrity of the original responses, no imputation was perceptions may reflect aggregated experiences across performed; instead, cases with missing data were excluded heterogeneous platforms. This approach aligns with the study’s listwise from the analysis. EFA was conducted using maximum objective of assessing trauma-informed characteristics of mental likelihood estimation on the correlation matrix, with an oblimin health chatbots at a conceptual level, while acknowledging that rotation applied to account for potential correlations among platform-level variability may influence factor structure and factors. The number of factors retained was guided by parallel outcome estimates. analysis. A stepwise item reduction procedure was applied, Overall, 62.4% of participants were female, 37.5% were aged sequentially removing the item with the lowest standardized 25–34 years, 75.2% identified as White, and 30.2% held a loading at each iteration until all retained items demonstrated bachelor’s degree. A majority reported a history of household primary factor loadings �0.50 to ensure strong associations with dysfunction (71.3%), followed by experiences of various forms of the extracted factors. Following item selection, the theoretical abuse. Detailed demographic characteristics of the final analytic framework was refereed to ensure retained items remained sample are presented in Table 1. conceptually coherent. CFA was subsequently conducted to test the fit of the TIC model identified in the EFA phase. CFA models were evaluated using standard goodness-of-fit indices, 2.4 Data analysis including the Comparative Fit Index (CFI), Tucker-Lewis Index (TLI), Root Mean Square Error of Approximation (RMSEA), Qualitative data from open-ended survey questions were and Standardized Root Mean Square Residual (SRMR). analyzed using inductive reflexive thematic analysis to identify Because EFA and CFA were performed on the same sample, recurring patterns and user-defined concepts related to trauma- results are exploratory and descriptive, intended to examine informed chatbot experiences. Reflexivity was maintained internal structure rather than definitive model validation. throughout by the first and last authors to ensure transparency Nevertheless, the combination of theory-informed item in interpretive decision-making [45]. development, stringent loading thresholds, oblique rotation, and Quantitative survey data, including multiple-response and multiple fit criteria supports the internal coherence and Likert-scale items, were analyzed using a combination of preliminary construct validity of the proposed TIC factor descriptive and inferential statistical methods. Descriptive structure. Future research should replicate and validate this statistics were first computed to summarize participant model in independent, clinically representative samples. characteristics, technology use patterns, and overall perceptions of trauma-informed care. Exploratory factor analysis (EFA) was 2.4.3 Logistic regression model then conducted to identify latent domains underlying participants’ conceptualization of trauma-informed chatbot Logistic regression analyses were conducted to examine characteristics. Confirmatory factor analysis (CFA) was predictors of whether participants perceived their chatbot subsequently performed to validate the factor structure and experience as trauma-informed. Composite scores were first evaluate model fit. Finally, multivariable logistic regression calculated from the validated CFA structure, with items loaded on models were fitted to identify factors associated with the same latent factor averaged to form summary scores for each participants’ perceptions of chatbot interactions as trauma- dimension. These composite scores were merged with informed. demographic variables and technology proficiency data to serve as All statistical analyses were conducted using R (version 4.4.3) predictors in a multivariable logistic regression model. A forward and relevant statistical packages. An overview of the analytic stepwise selection procedure was used to identify the most workflow used to address the study research questions is parsimonious model. Final model estimates are reported as odds presented in Figure 1. ratios (ORs) with corresponding p-values and 95% confidence Frontiers in Digital Health 04 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 TABLE 1 Demographic characteristics of participants by recruitment source (N = 606). Characteristics Clickworker Social media MyChart Overall (N = 488) (N = 113) (N = 5) (N = 606) Gender Female 326 (66.8%) 47 (41.6%) 5 (100.0%) 378 (62.4%) Male 152 (31.1%) 65 (57.5%) 0 (0.0%) 217 (35.8%) Nonbinary 8 (1.6%) 0 (0.0%) 0 (0.0%) 8 (1.3%) Transgender 1 (0.2%) 1 (0.9%) 0 (0.0%) 2 (0.3%) Prefer not to say 1 (0.2%) 0 (0.0%) 0 (0.0%) 1 (0.2%) Age group 18–24 years 37 (7.6%) 12 (10.6%) 0 (0.0%) 49 (8.1%) 25–34 years 174 (35.7%) 51 (45.1%) 2 (40.0%) 227 (37.5%) 35–44 years 183 (37.5%) 41 (36.3%) 1 (20.0%) 225 (37.1%) 45–54 years 87 (17.8%) 9 (8.0%) 2 (40.0%) 98 (16.2%) 55–64 years 7 (1.4%) 0 (0.0%) 0 (0.0%) 7 (1.2%) 65 years or older 0 (0.0%) 0 (0.0%) 0 (0.0%) 0 (0.0%) Race White 373 (76.4%) 79 (69.9%) 4 (80.0%) 456 (75.2%) Black 77 (15.8%) 23 (20.4%) 0 (0.0%) 100 (16.5%) Asian 37 (7.6%) 3 (2.7%) 0 (0.0%) 40 (6.6%) American Indian 15 (3.1%) 7 (6.2%) 0 (0.0%) 22 (3.6%) Pacific Islander 2 (0.4%) 2 (1.8%) 0 (0.0%) 4 (0.7%) Other 18 (3.7%) 0 (0.0%) 1 (20.0%) 19 (3.1%) Missing 1 (0.2%) 0 (0.0%) 0 (0.0%) 1 (0.2%) Adverse life experiences Abuse 323 (66.2%) 53 (46.9%) 5 (100.0%) 381 (62.9%) Adoption 19 (3.9%) 6 (5.3%) 1 (20.0%) 26 (4.3%) Child services involvement 79 (16.2%) 7 (6.2%) 2 (40.0%) 88 (14.5%) Community adversity 198 (40.6%) 39 (34.5%) 3 (60.0%) 240 (39.6%) Household dysfunction 372 (76.2%) 55 (48.7%) 5 (100.0%) 432 (71.3%) Neglect 234 (48.0%) 50 (44.2%) 4 (80.0%) 288 (47.5%) Education level Less than high school 12 (2.5%) 0 (0.0%) 0 (0.0%) 12 (2.0%) High school or GED 94 (19.3%) 3 (2.7%) 0 (0.0%) 97 (16.0%) Some college (no degree) 129 (26.4%) 5 (4.4%) 2 (40.0%) 136 (22.4%) Technical certification 25 (5.1%) 3 (2.7%) 0 (0.0%) 28 (4.6%) Associate degree 55 (11.3%) 26 (23.0%) 2 (40.0%) 83 (13.7%) Bachelor’s degree 127 (26.0%) 55 (48.7%) 1 (20.0%) 183 (30.2%) Master’s degree 41 (8.4%) 14 (12.4%) 0 (0.0%) 55 (9.1%) Doctoral degree 2 (0.4%) 7 (6.2%) 0 (0.0%) 9 (1.5%) Prefer not to say 2 (0.4%) 0 (0.0%) 0 (0.0%) 2 (0.3%) Missing 1 (0.2%) 0 (0.0%) 0 (0.0%) 1 (0.2%) Frontiers in Digital Health 05 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 FIGURE 1 Overview of the analytic workflow used to address the study research questions. intervals (CIs), providing insight into which chatbot features are MyChart participants (M = 4.10, SD = 0.71) had slightly lower most strongly associated with positive TIC perceptions. average scores across most domains. Despite these differences, mean scores remained high, indicating strong baseline digital literacy among participants. The details can be seen in 3 Result Supplementary Figure S1. 3.1 General descriptive statistics 3.1.3 Chatbot use and motivations 3.1.1 Overall satisfaction Participants reported varied engagement with mental health chatbots in terms of both frequency and platform type. The A total of 606 surveys were included in analysis. Overall, 92.9% results are presented in Figure 2. Most used the chatbot 1–3 of respondents perceived the chatbot as trauma-informed, with times per month (19.3%) or 2–5 times per week (19.0%), while similarly high endorsement across recruitment sources 17.5% reported trying it only once. Daily use was less common (Clickworker: 91.6%, social media: 100.0%, MyChart: 60.0%). (8.4%), indicating that chatbots may be serving as intermittent Likewise, 92.6% of participants reported they would recommend support tools rather than daily companions for most users. For the chatbot to others with a trauma history (Clickworker: 91.0%, platforms, the most frequently used chatbots were Serenity social media: 100.0%, MyChart: 80.0%). Satisfaction ratings (23.1%), Woebot (21.5%), and Stresscoach (19.8%). Participants further supported chatbot acceptability. On a five-point scale reported diverse motivations for engaging with mental health ranging from “very dissatisfied” to “very satisfied,” most chatbots. The most endorsed reasons were related to emotional participants expressed a positive experience, with 84.5% selecting distress, including stress or anxiety (68.8%) and depression the top two satisfaction levels (48.0% rated it as satisfied and (59.4%). A substantial proportion also reported exploratory 36.5% as extremely satisfied). Lower satisfaction ratings were rare, motives such as self-guided exploration (46.0%) and curiosity with only 2.9% of respondents expressing dissatisfaction. These (37.8%). The desire for a psychologically safe space was also a patterns were generally consistent across all recruitment sources. prevalent motivation (41.1%), ranking higher than expected clinical drivers such as trauma (32.5%) or loneliness (36.3%). 3.1.2 Technology experience Less commonly selected reasons included privacy concerns, recommendations, and previous experience, each under 20%. Participants demonstrated generally high levels of self-reported proficiency across all six technology domains: smartphone use, internet browsing, AI tools, computer use, social media, and smart home devices. Among these, smartphone use (M = 4.73, 3.2 RQ1: conceptualization of TIC SD = 0.54) and internet browsing (M = 4.71, SD = 0.56) principles proficiencies were rated highest across the full sample. In contrast, proficiency in smart home device use was the lowest overall (M = 3.2.1 Users perceptions across six core 3.99, SD = 1.02), reflecting broader variability in exposure to and dimensions use of emerging technologies. Clickworker respondents reported the highest average proficiency overall (M = 4.44, SD = 0.82), To explore how users conceptualized trauma-informed care followed by social media participants (M = 4.37, SD = 0.86). (TIC) in the context of mental health chatbots, participants’ Frontiers in Digital Health 06 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 FIGURE 2 Patterns of mental health chatbot use among participants. (A) Distribution of mental health applications used. (B) Self-reported frequency of application usage. (C) Primary reasons for using mental health applications. Respondents could select multiple options; percentages reflect the proportion of the total sample. responses to multiple-response checkbox items aligned with the frequently selected feature within this domain (86.7%), highlighting six core TIC principles were summarized. Both individual the importance of psychological security in chatbot interactions. The feature frequencies and co-occurrence patterns were examined most common feature combination (28.9%) included emotional to characterize how users collectively defined a trauma- safety, physical safety, and digital safety, indicating a holistic view of informed chatbot. safety spanning psychological and technical dimensions. Safety was primarily defined as the ability to interact with the Trust and transparency centered on ensuring that users feel chatbot without fear or judgment. Emotional safety was the most respected through open and clear communication. Information Frontiers in Digital Health 07 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 confidentiality emerged as the highest-priority feature (82.3%), participants’ responses aligned with the six theoretically specified underscoring strong concerns regarding privacy. A frequently TIC principles or reflected an alternative latent structure shaped occurring combination (26.7%) included transparency, tailored by the interactional affordances of chatbot systems. responses, validation of experiences, and data care, suggesting Parallel analysis supported a five-factor solution. The final that trust is closely linked to both personalization and EFA model retained 18 items and explained 66% of the total responsible data handling. variance, with individual factors accounting for 7% to 20% of Empowerment emphasized user autonomy and support. the explained variance. Based on item content and theoretical Guide conversation was the most frequently endorsed feature coherence, the five factors were labeled Peer Support, Trust and (73.1%), underscoring the importance of user-directed dialogue. Transparency, Data Safety, Empowerment, and Cultural The most common combination (20.8%) included sharing Sensitivity. These factors represent the most salient trauma- control, no forced choice, expression options, hopeful informed dimensions as experienced by users during real-world messaging, and sense of agency, reflecting preferences for tools chatbot interactions. that foster autonomy and emotional encouragement. Four of the five factors mapped closely to SAMHSA TIC Collaboration reflected partnership and shared decision- principles. Peer Support reflected users’ desire for connection to making. Respect choices was the most frequently selected design others with shared lived experiences, consistent with prior element (72.6%). A commonly endorsed combination (38.8%) findings that perceived social connectedness, whether direct or included goal setting, shared goals, and respect choices, mediated, is central to supportive digital mental health suggesting that participants valued mutuality and co-creation of experiences [47]. Trust and Transparency encompassed a therapeutic path. perceptions of reliability, expectation alignment, and Peer support focused on connection with others who have responsiveness, echoing HCI literature demonstrating that trust similar lived experiences. Ask for support was the most in conversational agents is shaped by predictability and clarity frequently selected feature (73.1%), highlighting the importance rather than anthropomorphism alone [48]. Empowerment of self-initiated outreach. The most common configuration captured feelings of agency, autonomy, and emotional self- (19.8%) included offer support, support available, relevant efficacy, aligning with self-determination theory and prior work connections, and virtual or in-person support, indicating on empowering AI design [49, 50]. Cultural Sensitivity reflected interest in accessible and meaningful peer engagement. inclusive language use and avoidance of triggering content, Cultural sensitivity emphasized inclusivity and respect for consistent with research highlighting linguistic safety as a key diverse identities. Nonjudgmental was the most frequently determinant of perceived care quality in digital mental health endorsed feature (79.4%), reflecting a strong desire for tools [51]. Data safety emerged as a distinct latent construct, acceptance. The most common feature combination (23.8%) encompassing items related to digital safety, privacy being included sensitive to identity, avoids offense, avoids triggers, respected, and careful handling of personal data, rather than as validates experience, shows interest, and culturally aligned, part of a broad, undifferentiated safety factor. highlighting preferences for chatbots that are affirming, The SAMHSA principle of Collaboration did not emerge as a inclusive, and mindful of cultural and individual differences. distinct factor. Instead, collaboration-related items loaded onto Full item-level response distributions are presented in adjacent domains, particularly Empowerment and Trust and Supplementary Figure S2, and detailed item wording is provided Transparency. This pattern suggests that users may conceptualize in Supplementary Table S2. collaboration with chatbots less as reciprocal partnership and more as feeling supported, respected, and enabled. Factor 3.2.2 Uncovering latent constructs of trauma- correlations ranged from 0.31 to 0.68, indicating related but informed experience distinct constructs, and mean item complexity was low (1.1), supporting factor interpretability. Model fit was acceptable Participants’ conceptualization of trauma-informed chatbot (RMSR = 0.02; x2 ¼ 146:61, df ¼ 73, p , :001), and internal design was further explored through Likert-scale ratings, in consistency was strong across factors (Cronbach’s which they indicated their level of agreement (strongly disagree a ¼ 0:77–0:92). Detailed factor loadings and communalities are to strongly agree) with statements reflecting the six TIC reported in Supplementary Table S3. principles. This measure captured the extent to which users endorsed statements such as, “My chatbot is sensitive to my 3.2.3 Validating the factor structure of trauma- concerns, identity, and/or historical experiences.” Overall, informed experience responses reflected favorable perceptions: 75.2% of participants selected a positive response (Agree: 35.5%; Strongly Agree: To evaluate the adequacy of the factor structure identified 39.7%), 16.7% were neutral, and only a small minority expressed through EFA, a CFA was conducted using maximum likelihood disagreement (Disagree: 5.6%; Strongly Disagree: 2.6%). Item- estimation. The five-factor, 18-item model demonstrated level response distributions are presented in Supplementary excellent fit to the data (CFI = 0.978; TLI = 0.973; RMSEA = Figure S3. 0.045, 90% CI: 0.037–0.052; SRMR = 0.033). All standardized While these descriptive findings suggest broadly positive user factor loadings were statistically significant (p , :001) and perceptions, prior HCI research cautions that high endorsement ranged from 0.70 to 0.88, indicating strong associations between alone does not reveal how users cognitively organize or observed indicators and their respective latent constructs. differentiate experiential constructs in human–AI interaction Among the five latent domains confirmed through CFA, Peer [46]. Accordingly, we conducted an EFA to examine whether Support emerged as a particularly cohesive and robust construct, Frontiers in Digital Health 08 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 with nearly all items demonstrating strong factor loadings. Most abstract TIC principles were enacted during chatbot interactions, notably, the item “Q42. Support Available: My mental health thereby reinforcing the quantitative findings. chatbot offers opportunities for peer support and community Within the Trust and Transparency domain, participants connections as part of my mental health management” exhibited consistently emphasized the importance of emotional validation the highest standardized loading in the entire model (0.88), and responsive communication. Users described feeling establishing it as a central indicator of this domain. understood by the chatbot, noting that it would “listen and Within the Trust and Transparency domain, the item “Q22. understand,” and that “I trust it more because it feels like I’m Meets Expectations: My chatbot responds to my experiences in a talking with a caring and concerned friend, not a chatbot that way that meets my expectations” had the highest loading (0.81), just gives basic random responses.” These accounts highlight highlighting its critical role in shaping user perceptions of how perceived empathy and tailored responses foster trustworthiness and responsiveness. In the Data Safety domain, trustworthiness in automated mental health support. “Q15. Privacy Respected: My mental health chatbot respects my Data Safety was described as a holistic construct privacy and confidentiality during interactions” demonstrated the encompassing anonymity, immediacy of access, and digital strongest loading (0.83), affirming that users prioritize digital protections. Several participants identified anonymity as central privacy and data protection as foundational to feeling safe in to their sense of safety, with one noting that “the main thing chatbot interactions. For Empowerment, the leading item was that helps me feel safe” was the ability to interact without “Q34. Sense of Agency: My chatbot helps me believe that I have revealing personal identity. Others emphasized the importance agency/ability to change things in my life,” which yielded a of data protection, explaining that “knowing my conversations loading of 0.88, underscoring the importance of fostering personal are secure and private makes me feel safer.” Together, these autonomy and emotional self-efficacy. Lastly, within Cultural responses illustrate how technical safeguards contribute to users’ Sensitivity, “Q51. Avoids Triggers: My chatbot avoids words or perceptions of safety. phrases that may trigger me” showed the strongest association Experiences related to Empowerment centered on agency, (0.81), emphasizing users’ sensitivity to emotionally safe language. encouragement, and goal support. Participants expressed Figure 3 presents a synthesized summary of findings, appreciation for features that promoted self-direction and integrating both the descriptive checkbox responses and the motivation, with one reporting that “consistent check-ins help CFA results. The detailed standardized loadings for each item keep me on track with my goals.” These reflections align with within the CFA model are presented in Figure 4 and the quantitative findings, underscoring the role of empowerment Supplementary Table S4, supporting the scale’s construct in shaping trauma-informed perceptions. validity and further clarifying item-factor relationships. Responses regarding Peer Support were more heterogeneous. While some participants preferred private, one-on-one chatbot interactions, others expressed interest in shared experiences, 3.2.4 Qualitative evidence supporting the five- noting that “peer-to-peer interactions. . . would be appreciated factor structure sometimes.” This variability suggests that peer support is valued selectively and may benefit from flexible, opt-in design approaches. Qualitative findings further supported the five-factor structure Finally, Cultural Sensitivity emerged as a critical dimension identified through EFA and CFA by illustrating how users of inclusion and respect. Participants highlighted the importance experienced each TIC domain in practice. Across domains, of consent, nonjudgmental language, and attentiveness to participants’ narratives provided concrete examples of how emotional boundaries, stating that “the bot does check in to see FIGURE 3 Summary of descriptive and confirmatory factor findings on trauma-informed chatbot features. The top portion summarizes participants’ most frequently endorsed features (circle endpoints) and commonly co-selected combinations (square endpoints) across the original six trauma- informed care principles. The bottom portion displays the final five-factor structure derived from confirmatory factor analysis, with darker colors indicating items with higher standardized factor loadings. Frontiers in Digital Health 09 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 FIGURE 4 Factor structure with standardized loadings for the Confirmatory Factor Analysis model of trauma-informed chatbot perception.TrS: Trust; Sft: Safety; Emp: Empowerment;PrS: Peer Support; SnS: Cultural Sensitivity. Bar widths represent the magnitude of standardized loadings, with values above 0.80 indicating particularly strong associations between items and their respective latent constructs. if I am comfortable” and “asks permission to discuss a sensitive 3.3.2 User characteristics topic further.” These accounts emphasize users’ expectations for culturally and emotionally attuned communication. Several user-level characteristics further differentiated trauma- informed perceptions. Participants who reported prior involvement with the Department of Child Services exhibited 3.3 RQ2: identifying predictors for chatbot markedly higher odds of perceiving the chatbot as trauma- being trauma-informed informed (OR = 9.71, 95% CI: 1.66–56.79, p = .012). Although confidence intervals were wide, reflecting limited subgroup size, To examine factors associated with perceiving a mental health this finding suggests that individuals with lived experience of chatbot as trauma-informed, we fitted a multivariable logistic formal care systems may be particularly attuned to trauma- regression model. Composite scores for the five CFA-derived aligned design features. dimensions were first entered into a baseline model. Forward In contrast, two predictors were negatively associated with stepwise selection was subsequently applied to identify a trauma-informed perception. Users who cited convenience as parsimonious set of predictors. Candidate variable reason: others their primary motivation for chatbot use had significantly lower were excluded from the final model due to estimation instability odds of perceiving the chatbot as trauma-informed (OR = 0.33, (e.g., near-infinite standard errors), driven by sparse category 95% CI: 0.11–0.96, p = .041). Similarly, higher self-rated endorsement and small subgroup sizes. The final model results smartphone proficiency was inversely associated with trauma- are summarized in Figure 5. informed perception (OR = 0.38, 95% CI: 0.16–0.93, p = .033). These associations may reflect differing evaluative standards 3.3.1 Chatbot design features among highly digitally proficient users, who may prioritize efficiency, performance, and usability over relational or Among design-related predictors, perceived Trust emerged as emotionally supportive features, and therefore be less responsive the strongest and most stable correlate of trauma-informed to trauma-informed design cues. perception. Higher trust scores were associated with substantially increased odds of perceiving the chatbot as trauma-informed (OR = 3.89, 95% CI: 1.73–8.77, p = .001). Peer Support (OR = 1.73, 4 Discussion 95% CI: 1.09–2.74, p = .021) and Empowerment (OR = 1.97, 95% CI: 1.09–3.57, p = .025) were also positively associated with Our study offers a rigorous empirical foundation for trauma-informed perception, reinforcing the importance of understanding how users interpret, experience, and evaluate TIC relational responsiveness and autonomy-supportive features in features in mental health chatbot interactions. By integrating shaping users’ evaluations of chatbot interactions. descriptive pattern analysis, confirmatory factor analysis, and Frontiers in Digital Health 10 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 FIGURE 5 Multivariable logistic regression results showing predictors of perceived trauma-informed chatbot interaction. Estimates are presented as odds ratios with 90% confidence intervals and corresponding p-values. predictive modeling, we elucidate both the conceptual structure of Together, these design elements illustrate that users interpret TIC as perceived by end-users and the key elements that drive trauma-informed chatbots through an integrated lens of trauma-informed appraisals. emotional security, autonomy, relational responsiveness, and Overall, users overwhelmingly perceived the chatbot as cultural respect. embodying TIC characteristics, with more than 90% endorsing Our factor analysis results yielded a five-factor structure their experience as trauma-informed alongside similarly high comprising Trust and Transparency, Data Safety, Empowerment, rates of satisfaction and recommendation. Importantly, Peer Support, and Cultural Sensitivity. The factor analysis participants did not interpret TIC as a monolithic construct. revealed a distinct Data Safety latent factor, which suggests that, Instead, both quantitative and qualitative findings revealed a for our population, safety was primarily understood and nuanced understanding spanning emotional safety, privacy, evaluated through the lens of data protection and privacy, rather identity validation, agency, and peer connection. These themes than a generalized sense of emotional or interpersonal safety. In align with existing TIC theory while contributing new empirical the context of mental health chatbots, this distinction indicates evidence to its application in digital mental health contexts [26, that assurances about how user data are stored, protected, and 27, 34, 42]. Rather than viewing TIC as a single construct, used may be especially salient for individuals with lived participants emphasized key design features that shaped how experiences of vulnerability or trauma and may meaningfully they judged the chatbot’s empathy, safety, and trustworthiness. shape their perceptions of whether an interaction feels safe. The Emotional Safety and Information Confidentiality were most absence of a distinct Collaboration factor, despite its theoretical valued, confirming that freedom from fear, judgment, and salience, suggests that collaborative elements (e.g., shared privacy concerns defines psychological safety in digital care [2, decision-making) may be implicitly embedded within other 19, 20, 52]. Features that allowed users to guide conversations domains. Qualitative accounts reinforced this interpretation: and receive hopeful, personalized responses fostered several participants emphasized valuing “consistent feedback” empowerment and a sense of agency [52, 53]. Respect for user and the sense that responses were “tailored to me,” choices and adaptive feedback conveyed collaboration and underscoring collaboration as an integrated rather than mutuality without explicit co-decision-making [29, 54, 55]. standalone dimension. Opportunities for self-initiated peer connection further Trust and Transparency emerged as the most robust driver of enhanced perceived support and belonging [47, 56]. Finally, a trauma-informed perception, especially in relation to nonjudgmental tone and cultural awareness reinforced confidentiality and expectation management. Emotional inclusivity and emotional comfort across backgrounds [39]. validation played a central role in cultivating trust, as reflected Frontiers in Digital Health 11 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 in both model estimates and user narratives. Participants or less threatening than human interlocutors, in part because consistently described the importance of feeling “listened to and they offer predictable, nonjudgmental interactions and reduce understood” and noted that empathetic language made them fears of stigma or social evaluation [61, 62, 72]. Youth in foster feel “like I’m talking with a caring and concerned friend, not care often report mistrust of service providers and concerns just a chatbot.” Such validation highlights how relational design about stigma when seeking mental health support , barriers that elements underpin trust and drive trauma-informed evaluations. may be attenuated in interactions with digital agents [61, 63, This finding aligns with prior research demonstrating that 72]. Similar findings have been reported in digital mental health relational authenticity and transparent communication are contexts, where users with greater distress often report higher essential for trust-building in digital mental health tools [19, 20]. perceived empathy and trust in chatbot-based support tools [30, Empowerment and Peer Support also proved to be critical 73]. Moreover, research specifically examining technology use dimensions for shaping trauma-informed perceptions. Users among foster care populations indicates that digital platforms who felt supported in decision-making or had access to may offer a sense of privacy and agency that is particularly community-based resources were significantly more likely to valued by those who have experienced institutional oversight view the chatbot as trauma-informed, echoing findings by and limited control over their lives [57, 60, 74]. Potts et al., who observed that multilingual chatbots offering Conversely, users who cited convenience as their primary user-driven dialogue improved well-being and engagement reason for chatbot use, or who rated themselves highly [53]. Similarly, participants valued access to supportive, proficient with smartphones, were significantly less likely to community-oriented features, paralleling Beilharz et al., who perceive the chatbot as trauma-informed. This pattern aligns emphasized that peer-based and family-linked chatbot with prior work suggesting that technologically experienced or components enhanced feelings of connection and emotional efficiency-oriented users tend to evaluate conversational systems safety among adolescents [56]. This aligns with core TIC through a utilitarian lens, prioritizing speed, task completion, principles that emphasize relational healing and shared agency and functional accuracy over relational or emotional qualities as foundational to care [4, 27]. [75, 76]. Additionally, higher smartphone proficiency may foster Individual differences further shaped TIC perceptions. greater critical awareness of automated empathy, leading some Participants in our study with a history of involvement with users to approach relational or caring language with increased child protective or social services had significantly higher odds skepticism [77]. Related work in digital self-care and health of perceiving the chatbot as trauma-informed.This finding may technologies has similarly shown that users motivated by be understood through the lens of system-impacted populations’ convenience often place lower value on emotional depth, which unique relationships with institutional support and technology. can lead to reduced perceptions of empathy or trust [54]. In Research on youth with foster care or child welfare involvement such cases, trauma-informed features, such as validation, has documented that such experiences often lead to distinctive reflective language, or user control, may be perceived as patterns of help-seeking behavior and trust formation [57–60]. peripheral or insufficient relative to expectations for Youth in foster care frequently experience multiple placements seamless performance. and disrupted relationships, which can profoundly shape their Overall, our results offer critical insights for the future of expectations and evaluations of new support systems, whether trauma-informed AI in mental health. First, we demonstrate human or technological [58, 60–63]. Studies examining digital that TIC is both legible and empirically measurable: users health adoption among foster youth suggest that this population overwhelmingly endorsed trauma-informed experiences, and approaches technology-mediated support with both strategic factor analysis revealed a robust five-domain structure that caution and openness, particularly valuing features that signal provides strong evidence for measurement accuracy. Second, safety, autonomy, and non-coercive care [59, 60, 64]. These predictive modeling identified high-impact predictors of trauma- individuals often develop what might be termed “system informed perceptions, spanning both chatbot design features literacy”: a heightened ability to discern whether systems, (e.g., trust, empowerment, and peer support) and user including digital ones, are designed with their safety and dignity characteristics (e.g., trauma history, motivations for use, and in mind [61–63]. Prior work has also shown that young people technology proficiency), thereby underscoring the importance of who have experienced foster care, adoption, or removal from personalization and user-centered design. Third, our findings biological families may be particularly attuned to relational reveal meaningful variation in how different user populations qualities in both human and technological interactions, given perceive and prioritize TIC features, suggesting that one-size- their histories of disrupted attachments and mandated fits-all approaches may inadequately serve diverse trauma institutional relationships [58, 65–69]. survivors. Designers of mental health chatbots should embed Consistent with this interpretation, research on relational trauma-informed principles not as optional enhancements but agents has shown that users experiencing higher emotional as core architectural features. Likewise, TIC-aligned metrics vulnerability, including depressive symptoms or psychological should be systematically incorporated into evaluation distress, tend to form stronger therapeutic alliances with frameworks. As reliance on automated tools for mental health conversational systems [70, 71] For individuals with child support continues to grow, establishing rigorous standards for welfare histories, this effect may be compounded: digital agents safety, ethics, and human-centered care will be essential. may represent a form of support perceived as less surveilling, Several limitations should be considered when interpreting less coercive, and more user-controlled than traditional these findings. First, the study employed a cross-sectional, self- institutional services [57, 61, 63, 72]. These studies suggest that reported survey design, which may be influenced by recall bias conversational agents may be experienced as emotionally safer or social desirability effects. Although the sample was Frontiers in Digital Health 12 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 demographically diverse, it was predominantly White and limited Written informed consent for participation in this study was to English-speaking individuals with reliable internet access, obtained directly from all participants prior to their involvement potentially restricting generalizability to more diverse or digitally in the study. marginalized populations. Future research should intentionally oversample underrepresented groups, including racial and ethnic minorities, non-English speakers, and individuals with limited Author contributions digital access, to ensure equity in TIC design and evaluation [78]. Second, despite multiple attempts to recruit through the XL: Formal analysis, Methodology, Visualization, Writing – MyChart patient portal, only five participants were enrolled via original draft. SG: Conceptualization, Methodology, Writing – this channel, limiting meaningful comparisons across review & editing. RP: Data curation, Writing – review & editing. recruitment sources and underrepresenting clinical populations. JP: Writing – review & editing. CW: Conceptualization, Third, while the regression results highlight meaningful Writing – review & editing. FN: Conceptualization, Data associations between trauma-informed perception and both curation, Methodology, Writing – review & editing. chatbot design features and user characteristics, the high prevalence of positive outcomes introduces ceiling effects that may inflate odds ratios and reduce estimate stability for rare subgroups. Accordingly, these findings should be interpreted as Funding identifying salient patterns and hypotheses for future validation in more balanced and clinically diverse samples. Finally, The author(s) declared financial support was received for this although the study relied primarily on quantitative survey data, work and/or its publication. This work was conducted as part of a which provides breadth, it offers limited depth regarding lived study supported by the National Science Foundation Award No. experiences. To strengthen data triangulation, the research team 2348691. is conducting complementary semi-structured interviews as part of the broader project. In conclusion, our study establishes a rigorous empirical Conflict of interest foundation for understanding how TIC principles manifest in mental health chatbot interactions. By integrating descriptive The author(s) declared that this work was conducted in the analysis, factor modeling, predictive regression, and qualitative absence of any commercial or financial relationships that could insights, we demonstrate that TIC is both measurable and be construed as a potential conflict of interest. impactful, with trust, empowerment, and peer support emerging as central drivers of user perceptions alongside key individual characteristics. To advance trauma-informed digital mental Generative AI statement health, developers should prioritize hybrid models that integrate peer networks and human–bot collaboration, tailor functionality The author(s) declared that Generative AI was not used in the to diverse user profiles, and systematically evaluate emerging creation of this manuscript. TIC dimensions. Future longitudinal and multi-platform Any alternative text (alt text) provided alongside figures in this research will be critical for validating and extending these article has been generated by Frontiers with the support of findings, ultimately guiding the design of more inclusive, artificial intelligence and reasonable efforts have been made to effective, and psychologically safe chatbot interventions. As ensure accuracy, including review by the authors wherever artificial intelligence becomes increasingly embedded in mental possible. If you identify any issues, please contact us. health service delivery, the ethical imperative to center trauma survivors’ voices, protect their safety, and honor their agency has never been more urgent [79]. This study represents an essential step toward realizing that vision. Publisher’s note All claims expressed in this article are solely those of the Data availability statement authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the The original contributions presented in the study are publicly reviewers. Any product that may be evaluated in this article, or available. This data can be found here: https://gitlab.com/ claim that may be made by its manufacturer, is not guaranteed E190153/user-perceptions-of-trauma-informed-chatbots. or endorsed by the publisher. Ethics statement Supplementary material The studies involving humans were approved by the Parkview The Supplementary Material for this article can be found Health Institutional Review Board. The studies were conducted in online at: https://www.frontiersin.org/articles/10.3389/fdgth.2026. accordance with local legislation and institutional requirements. 1797681/full#supplementary-material Frontiers in Digital Health 13 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 References 1. Kulkarni P, Mahabaleshwarkar A, Kulkarni M, Sirsikar N, Gadgil KD. 23. Xiang C. Data from: “He would still be here”: man dies by suicide after talking Conversational AI: an overview of methodologies, applications & future scope. In: with AI chatbot, widow says. Vice (2023). https://www.vice.com/en/article/man- Proceedings of the International Conference on Computing, Communication, Control dies-by-suicide-after-talking-with-ai-chatbot-widow-says (Accessed March 30). and Automation (ICCUBEA). Pune, India: IEEE (2019). p. 1–6. doi: 10.1109/ 24. Eggleston M, Jones EP, Khan N, Romani WA, McQuillan K, Otero J, et al. A ICCUBEA47591.2019.9129347 scoping review of trauma-informed care principles applied in design and 2. Fitzpatrick KK, Darcy A, Vierhile M. Delivering cognitive behavior therapy to technology. Digit Health. (2025) 11. doi: 10.1177/20552076251360925 young adults with symptoms of depression and anxiety using a fully automated 25. Mazuz K, Yamazaki R. Trauma-informed care approach in developing conversational agent (woebot): a randomized controlled trial. JMIR Ment Health. companion robots: a preliminary observational study. Front. Rob AI. (2025) (2017) 4:e19. doi: 10.2196/mental.7785 12:1476063. doi: 10.3389/frobt.2025.1476063 3. Gaffney H, Mansell W, Tai S. Conversational agents in the treatment of mental 26. Substance Abuse and Mental Health Services Administration. Trauma-informed health problems: mixed-method systematic review. JMIR Ment Health. (2019) 6: care in behavioral health services. No. HHS Publication No. (SMA) 13-4801 in e14166. doi: 10.2196/14166 Treatment Improvement Protocol (TIP) Series 57. Rockville, MD: Substance Abuse 4. Abd-Alrazaq AA, Rababeh A, Alajlani M, Bewick BM, Househ M. Effectiveness and Mental Health Services Administration (2014b). (Accessed January 20, 2026). and safety of using chatbots to improve mental health: systematic review and meta- 27. Substance Abuse and Mental Health Services Administration. SAMHSA’s concept analysis. J Med Internet Res. (2020) 22:e16021. doi: 10.2196/16021 of trauma and guidance for a trauma-informed approach (2014a). Tech. rep., 5. Han Q, Zhao C. Unleashing the potential of chatbots in mental health: bibliometric SAMHSA, Rockville, MD. HHS Publication No. SMA14-4884. analysis. Front Psychiatry. (2025) 16:1494355. doi: 10.3389/fpsyt.2025.1494355 28. Abdulai A-F. Is generative AI increasing the risk for technology-mediated trauma 6. Li H, Zhang R, Lee YC, Kraut RE, Mohr DC. Systematic review and meta-analysis among vulnerable populations? Nurs Inq. (2025) 32:e12686. doi: 10.1111/nin.12686 of ai-based conversational agents for promoting mental health and well-being. NPJ 29. Chen JX, McDonald A, Zou X, Tseng E, Roundy KA, Tamersoy A, et al. Trauma- Digit Med. (2023) 6:1–14. doi: 10.1038/s41746-023-00979-5 informed computing: towards safer technology experiences for all. In: Proceedings of 7. Meheli S, Sinha C, Kadaba M. Understanding people with chronic pain who use a the CHI Conference on Human Factors in Computing Systems. ACM (2022). p. 1–15. cognitive behavioral therapy–based artificial intelligence mental health app (Wysa): doi: 10.1145/3491102.3517475 mixed methods retrospective observational study. JMIR Hum Factors. (2022) 9: 30. Felitti VJ, Anda RF, Nordenberg D, Williamson DF, Spitz AM, Edwards V, et al. e35671. doi: 10.2196/35671 Relationship of childhood abuse and household dysfunction to many of the leading 8. Haque R, Rubya S. An overview of chatbot-based mobile mental health causes of death in adults: the adverse childhood experiences (ACE) study. Am applications: insights from app description and user reviews. JMIR Mhealth J Prev Med. (1998) 56:774–86. doi: 10.1016/S0749-3797(98)00017-8 Uhealth. (2023) 11:e44838. doi: 10.2196/44838 31. Gerber MR, Elisseou S, Sager ZS, Keith JA. Trauma-informed telehealth in the 9. Lau N, O’Daffer A, Colt S, Yi-Frazier JP, Palermo TM, McCauley E, et al. Android COVID-19 era and beyond. Fed Pract. (2020) 37:302–8. doi: 10.1016/S0749-3797 and iphone mobile apps for psychosocial wellness and stress management: systematic (98)00017-8 search in app stores and literature review. JMIR Mhealth Uhealth. (2020) 8:e17798. 32. Muskett C. Trauma-informed care in inpatient mental health settings: a review of doi: 10.2196/17798 the literature. Int J Ment Health Nurs. (2014) 23:51–9. doi: 10.1111/inm.12012 10. Vaidyam AN, Wisniewski H, Halamka JD, Kashavan MS, Torous JB. Chatbots 33. Randazzo C, Scott CF, Bellini R, Ammari T, Devito MA, Semaan B, et al. Trauma- and conversational agents in mental health: a review of the psychiatric landscape. informed design: a collaborative approach to building safer online spaces. In: Proceedings Can J Psychiatry. (2019) 64:456–464. doi: 10.1177/0706743719828977 Companion Publication of the Conference on Computer Supported Cooperative Work and 11. Farzan M, Ebrahimi H, Pourali M, Sabeti F. Artificial intelligence-powered Social Computing. ACM (2023). p. 203–7. doi: 10.1145/3584931.3611277 cognitive behavioral therapy chatbots, a systematic review. Iran J Psychiatry. (2025) 34. Substance Abuse and Mental Health Services Administration. Data from: 20:102–10. doi: 10.18502/ijps.v20i1.17395 Trauma-informed approaches and programs (2024). Available online at: https:// 12. Gaffney H, Mansell W, Edwards R, Wright J. Manage your life online (MYLO): a www.samhsa.gov/mental-health/trauma-violence/trauma-informed-approaches- pilot trial of a conversational computer-based intervention for problem solving in a programs (Accessed December 3, 2024). student sample. Behav Cogn Psychother. (2014) 42:731–46. doi: 10.1017/ 35. Lee HS, Wright C, Ferranto J, Buttimer J, Palmer CE, Welchman A, et al. Artificial S135246581300060X intelligence conversational agents in mental health: patients see potential, but prefer 13. Denecke K, Gabarron E. The ethical aspects of integrating sentiment and emotion humans in the loop. Front Psychiatry. (2025) 15:1505024. doi: 10.3389/fpsyt.2024. analysis in chatbots for depression intervention. Front Psychiatry. (2024) 15:1462083. 1505024 doi: 10.3389/fpsyt.2024.1462083 36. Schäfer LM, Krause T, Köhler S. Exploring user characteristics, motives, and 14. Hipgrave L, Goldie J, Dennis S, Coleman A. Balancing risks and benefits: expectations and the therapeutic alliance in the mental health conversational AI clinicians’ perspectives on the use of generative AI chatbots in mental healthcare. Clare®: a baseline study. Front Digit Health. (2025) 7:1576135. doi: 10.3389/fdgth. Front Digit Health. (2025) 7:1606291. doi: 10.3389/fdgth.2025.1606291 2025.1576135 15. Lawrence HR, Schneider RA, Rubin SB, Matarić MJ, McDuff DJ, Jones Bell M. 37. Zou X, Na Y, Lai K, Liu G. Unpacking public resistance to health chatbots: a The opportunities and risks of large language models in mental health. JMIR Ment parallel mediation analysis. Front Psychol. (2024) 15:1276968. doi: 10.3389/fpsyg. Health. (2024) 11:e59479. doi: 10.2196/59479 2024.1276968 38. Kollig F, Voelker K, Ryan E, Pfafman R, Nova FF. Current landscape of mental 16. Rahsepar Meadi M, Sillekens T, Metselaar S, van Balkom A, Bernstein J, Batelaan health conversational agents from a trauma-informed care lens: scoping review N. Exploring the ethical challenges of conversational AI in mental health care: (2025). JMIR Preprints [Accepted preprint]. scoping review. JMIR Ment Health. (2025) 12:e60432. doi: 10.2196/60432 39. Kostenius C, Lindström F, Potts C, Pekkari N. Young peoples’ reflections about 17. Abd-Alrazaq A, Al-Mudawar M, Alhuwail D, Al-Kuwari S, Nashwan AJ, Shah Z. using a chatbot to promote their mental wellbeing in northern periphery areas—a Large language models in mental health care: a scoping review. npj Digit Med. (2023) qualitative study. Int J Circumpolar Health. (2024) 83. doi: 10.1080/22423982.2024. 6:167. doi: 10.1038/s41746-023-00898-7 2369349 18. Thirunavukarasu AJ, Ting DSJ, Elangovan K, Gutierrez L, Tan TF, Ting DSW. 40. National Coalition for Sexual Health. Data from: Trauma-informed care guide Large language models in medicine. Nat Med. (2023) 29:1930–40. doi: 10.1038/ (2023). Available online at: https://www.nationalcoalitionforsexualhealth.org/tools/ s41591-023-02448-8 for-healthcare-providers/trauma-informed-care-guide (Accessed March 31, 2026). 19. De Freitas J, Uğuralp AK, Oğuz-Uğuralp Z, Puntoni S. Chatbots and mental 41. Scott CF, Marcu G, Anderson RE, Newman MW, Schoenebeck S. Trauma- health: insights into the safety of generative AI. J Consum Psychol. (2024) informed social media: towards solutions for reducing and healing online harm. 34:481–91. doi: 10.1002/jcpy.1393 In: Proceedings of the 2023 CHI Conference on Human Factors in Computing 20. Scholich T, Barr M, Wiltsey Stirman S, Raj S. A comparison of responses from Systems. New York, NY, USA: Association for Computing Machinery (2023). CHI human therapists and large language model–based chatbots to assess therapeutic ’23. doi: 10.1145/3544548.3581512 communication: mixed methods study. JMIR Ment Health. (2025) 12:e69709. 42. Substance Abuse and Mental Health Services Administration. Practical guide for doi: 10.2196/69709 implementing a trauma-informed approach (2023). Tech. rep., SAMHSA, Rockville, 21. Brown JEH, Halpern J. AI chatbots cannot replace human interactions in the MD. HHS Publication No. PEP23-06-05-005. pursuit of more inclusive mental healthcare. SSM Ment Health. (2021) 1:100017. 43. Clickworker GmbH. Data from: Clickworker: crowdsourcing marketplace (2026). doi: 10.1016/j.ssmmh.2021.100017 Available online at: https://www.clickworker.com/ (Accessed January 14, 2026). 22. Roose K. Data from: Can AI be blamed for a teen’s suicide? The New York Times. 44. Gnewuch U, Morana S, Hinz O, Kellner R, Maedche A. More than a bot? The (2024). Available online at: https://www.nytimes.com/2024/10/23/technology/ impact of disclosing human involvement on customer interactions with hybrid characterai-lawsuit-teen-suicide.html (Accessed October 23). service agents. Inf Syst Res. (2024) 35:936–55. doi: 10.1287/isre.2022.0152 Frontiers in Digital Health 14 frontiersin.org Lu et al. 10.3389/fdgth.2026.1797681 45. Braun V, Clarke V. Reflecting on reflexive thematic analysis. Qual Res Sport Exerc 63. Singer ER, Berzin SC. Former foster system youth: perspectives on transitional Health. (2019) 11:589–97. doi: 10.1080/2159676X.2019.1628806 supports and programs. J Soc Serv Res. (2020) 48:287–305. 46. Luger E, Sellen A. “Like having a really bad PA”: the gulf between user expectation 64. Leo HP, Folk JB, Rodriguez C, Tolou-Shams M. Implementation considerations and experience of conversational agents. In: Proceedings of the 2016 CHI Conference for family-based telehealth interventions for youth in foster care: focus group study on Human Factors in Computing Systems. (2016a). p. 5286–97. doi: 10.1145/2858036. with child welfare system professionals. JMIR Form Res. (2023) 7:e45905. doi: 10. 2858288 2196/45905 47. Naslund JA, Aschbrenner KA, Marsch LA, Bartels SJ. The future of mental health 65. Advokids. Data from: Attachment (2024). Available online at: https://advokids. care: peer-to-peer support and social media. Epidemiol Psychiatr Sci. (2016) org/childhood-mental-health/attachment/ (Accessed December 13, 2025). 25:113–22. doi: 10.1017/S2045796015001067 66. Bovenschen I, Lang K, Zimmermann J, Förthner J, Nowacki K, Roland I, et al. 48. Seeger A-M, Pfeiffer J, Heinzl A. Texting with humanlike conversational agents: Foster children’s attachment security in the first year after placement: a designing for anthropomorphism. J Assoc Inf Syst. (2021) 22:931–67. doi: 10.17705/ longitudinal study of predictors. Early Child Res Q. (2016) 36:269–80. doi: 10.1016/ 1jais.00685 S0749-3797(98)00017-8 49. Calvo R, Peters D, Vold K, Ryan R. Supporting human autonomy in AI systems: a 67. Considering Adoption. Data from: Attachment in fostering and adoption (2020). framework for ethical enquiry. In: Philosophical Studies Series. Cham: Springer Available online at: https://consideringadoption.com/parenting-after-foster-care- (2020). p. 31–54. doi: 10.1007/978-3-030-50585-1_2 adoption/rad-and-attachment/ (Accessed December 13, 2025). 50. Ryan RM, Deci EL. Self-Determination Theory: Basic Psychological Needs in 68. Indiana Adoption Lawyer. Data from: Summary of attachment and bonding in the Motivation, Development, and Wellness. New York, NY: Guilford Press (2017). foster and adopted child (2024). Available online at: https://www. 51. Bickmore TW, Trinh H, Olafsson S, O’Leary TK, Asadi R, Rickles NM, et al. hoosierfamilylawyer.com/foster-children-adoption-book/summary-attachment-and- Patient and consumer safety risks when using conversational assistants for medical bonding-foster-and-adopted-child (Accessed December 14, 2025). information: an observational study of Siri, Alexa, and Google assistant. J Med 69. Research in Practice. Data from: Attachment theory and research: fostering and Internet Res. (2018) 20:e11510. doi: 10.2196/11510 adoption (2024). Available online at: https://fosteringandadoption.rip.org.uk/topics/ 52. He Y, Yang L, Zhu X, Wu B, Zhang S, Qian C, et al. Mental health chatbot for attachment-theory-research/ (Accessed December 14, 2025). young adults with depressive symptoms during the COVID-19 pandemic: single- 70. Bickmore TW, Mitchell SE, Jack BW, Paasche-Orlow MK, Pfeifer LM, O’Donnell blind, three-arm randomized controlled trial. J Med Internet Res. (2022) 24:e40719. J. Response to a relational agent by hospital patients with depressive symptoms. doi: 10.2196/40719 Interact Comput. (2010) 22:289–98. doi: 10.1016/j.intcom.2009.12.001 53. Potts C, Lindström F, Bond R, Mulvenna M, Booth F, Ennis E, et al. A multilingual 71. Bickmore TW, Picard RW. Establishing and maintaining long-term human- digital mental health and wellbeing chatbot (chatpal): pre-post multicenter intervention computer relationships. ACM Trans Comput Hum Interact. (2005) 12:293–327. study. J Med Internet Res. (2023) 25:e43051. doi: 10.2196/43051 doi: 10.1145/1067860.1067867 54. Moilanen J, van Berkel N, Visuri A, Gadiraju U, van der Maden W, Hosio S. 72. Moussavi Y, Haugland BSM, Wergeland GJ, Lehmann S, Mæland S. Experiences Supporting mental health self-care discovery through a chatbot. Front Digit Health. from treatment for anxiety and depression among youth in foster care: a qualitative (2023) 5:1034724. doi: 10.3389/fdgth.2023.1034724 study. Child Fam Soc Work. (2025) 30:342–53. doi: 10.1111/cfs.13172 55. Tremain H, McEnery C, Fletcher K, Murray G. The therapeutic alliance in digital 73. Inkster B, Sarda S, Subramanian V. An empathy-driven, conversational mental health interventions for serious mental illnesses: narrative review. JMIR Ment artificial intelligence agent (Wysa) for digital mental well-being: Real-world Health. (2020) 7:e17204. doi: 10.2196/17204 data evaluation mixed-methods study. JMIR Mhealth Uhealth. (2018) 6:e12106. 56. Beilharz F, Sukunesan S, Rossell SL, Kulkarni J, Sharp G. Development of a doi: 10.2196/12106 positive body image chatbot (kit) with young people and parents/carers: qualitative 74. iFoster. Data from: Foster youth are the most disconnected young people (2024). focus group study. J Med Internet Res. (2021) 23:e27807. doi: 10.2196/27807 Available online at: https://www.ifoster.org/bridging-the-digital-divide/ (Accessed 57. Ahlers K, Gabrielli J. A systematic review of internet communication technology December 14, 2025). use by youth in foster care. Child Adolesc Social Work J. (2021) 39:375–90. 75. Følstad A, Nordheim CB, Bjørkli CA. What makes users trust a chatbot for 58. Brown D. Childhood experiences, growing up “in care,” and trust: a quantitative customer service? An exploratory interview study. In: Bodrunova SS. editor. analysis. Child Youth Serv Rev. (2022) 140:106584. doi: 10.1016/j.childyouth.2022. Internet Science: 5th International Conference, INSCI 2018. Cham: Springer (2018). 106584 p. 194–208. Lecture Notes in Computer Science; Vol. 11193. doi: 10.1007/978-3- 030-01437-7_16 59. Folk JB, Gill AM, Ordorica C, Rodriguez C, Holloway IW, Meza RD, et al. An ecodevelopmental framework for engaging diverse youth in foster care and their 76. Luger E, Sellen A. “Like having a really bad PA”: the gulf between user expectation families into technology-based family intervention research trials. Front Digit and experience of conversational agents. In Proceedings of the 2016 CHI Conference Health. (2022) 4:866139. doi: 10.1016/S0749-3797(98)00017-8 on Human Factors in Computing Systems. New York, NY, USA: ACM (2016b). CHI ’16. p. 5286–97. doi: 10.1145/2858036.2858288 60. Patmisari E, Jones M, Kang J, Metelak R, Ordorica C, Tolou-Shams M. Reflections of foster youth engaging in the co-design of digital mental health technology: 77. Volpato R, DeBruine L, Stumpf S. Trusting emotional support from generative duoethnography study. JMIR Ment Health. (2025) 9:e53231. doi: 10.1016/S0749- artificial intelligence: a conceptual review. Comput Hum Behav Artif Hum. (2025) 3797(98)00017-8 5:100195. doi: 10.1016/j.chbah.2025.100195 61. Ahrens KR, Garnett MF, Dugan J, McCarthy M, Doty JG, Schultz S, et al. 78. Veinot TC, Mitchell H, Ancker JS. Good intentions are not enough: How Intervention development to improve foster youth mental health by targeting informatics interventions can worsen inequality. J Am Med Inform Assoc. (2018) coping self-efficacy and help-seeking. J Child Fam Stud. (2023) 144:106753. 25:1080–8. doi: 10.1093/jamia/ocy052 62. Bennett K, Courtney J, Decker K, Duppong Hurley K, Pfeiffer H. Trust is 79. Martínez-Martín N, Luo Z, Kaushal A, Adeli E, Haque A, Kelly S, et al. Ethical essential: identifying trust building techniques from youth providers across the issues for direct-to-consumer digital psychotherapy apps: addressing service array. Child Youth Serv Rev. (2020) 116:105320. doi: 10.1016/j.childyouth. accountability, data protection, and consent. JMIR Ment Health. (2020) 7:e15423. 2020.105320 doi: 10.2196/15423 Frontiers in Digital Health 15 frontiersin.org