Participants were recruited from a public university in New Jersey. Students learned about the study through an email newsletter that lists on-campus research opportunities that were sent to all students. The email provided information about the study and a link to a brief web-based survey to assess eligibility. After providing informed consent, participants were directed to take the screening survey to determine eligibility for the study. To be included, participants need to: (1) be aged 18 years or older; (2) be a student enrolled in the participating university; (3) live in the United States; (4) read and understand English; and (5) have elevated depression (Patient Health Questionnaire-2 [PHQ-2] [15] score ≥3) or anxiety symptoms (Generalized Anxiety Disorder 2 [GAD-2; [16] score ≥3).

Eligible participants were asked to complete a preintervention survey and to spend at least 5 minutes during their first Wayhaven conversation. At the end of the first session, they were directed to complete a postintervention survey. Following this initial conversation, participants continued to have access to Wayhaven for 1 week and were encouraged to engage with the AI mental wellness chatbot as much as needed. Participants received a final follow-up survey 1 week after enrollment. Each survey took 30‐45 minutes to complete. Participants were recruited from September 20 to September 25, 2024. Recruitment was closed once 40 students completed the study (ie, all assessments and at least one conversation). After completing the study, students were granted unlimited access to Wayhaven.

The Montclair State University Institutional Review Board approved this study (IRB-FY24-25-3875). All participants provided informed consent prior to particpating and were informed that they could to opt out of the study at any time. Participants received a US $20 gift card for completing each survey. All participant data was de-identified.

Participants were given access to Wayhaven upon study enrollment. Wayhaven is a generative AI-powered mental wellness chatbot, designed to deliver brief, evidence-based text conversations that support college students’ mental wellness needs. These conversations are rooted in CBT, dialectical behavior therapy, acceptance and commitment therapy, and mindfulness-based practices, developed in collaboration with mental health experts, including psychologists and mental health counselors. Student feedback has also been incorporated into Wayhaven’s working model. Specifically, before initiating this study, 74 students across 8 college campuses tested Wayhaven and responded to a brief open-ended survey eliciting feedback on the app. Students were also asked to participate in 45-minute focus groups (3‐6 students per group) where they completed design thinking sessions related to a particular feature or conversation flow change in the app. Six focus groups were conducted prior to the study.

Wayhaven aims to provide comprehensive mental wellness support to college students by offering tailored psychoeducational content and emotional support resources. The AI is designed to replicate empathic responses and offer specific techniques to address emotions or concerns shared by users. For example, a user expressing stress might first be validated and supported by the coach’s comments and then guided through a relaxation technique, ensuring a responsive and meaningful experience. Before starting a conversation, users select a coach persona that aligns with their preferences, such as a first-generation Latinx student or a university professor. This user-centered approach enhances engagement and flexibility, allowing users to switch between coaches while maintaining comprehensive support. Wayhaven also integrates university-specific context, such as details about campus resources, student clubs, and upcoming events, to ensure users have quick access to services available on their campus. By combining evidence-based practices with personalized AI interactions, Wayhaven not only addresses mental wellness concerns but also amplifies the utility of a school’s existing support infrastructure.

The user journey begins with account creation and onboarding, where participants provide demographic information and choose an AI mental wellness coach. Participants are able to choose from four options: Marcela (Latinx college student), Michelle (Black life coach), Professor Wilson (White college professor), and Coach Thompson (Black sports coach). Conversations follow a structured single-session intervention format, starting with the identification of a concern and a primary goal. The AI mental wellness chatbot then offers evidence-based tools, helps the user practice selected techniques, and collaborates to create an actionable plan. Each session concludes with a summary, and users can return for on-demand support at any time. To promote continued engagement, Wayhaven also sends push notifications to check in on users, ensuring ongoing and adaptive support. Figure 1 shows example conversations with the AI mental wellness coaches.

When creating an account, users are also informed that whenever they use Wayhaven, they will be chatting with an AI and not a real person, and that Wayhaven is not designed to address crises or substitute professional mental health care. Users seeking mental health services or reporting elevated distress are directed to campus-specific resources, such as the college counseling center. The app also features an “SOS” button, which allows users to connect to the Suicide and Crisis Lifeline (988) or the Crisis Text Line directly from the app. This feature appears at the top of each screen a user sees.

Analyses were conducted using SPSS (version 28; IBM Corp). Descriptive statistics were used to summarize participant characteristics, primary outcomes, app use and engagement, and satisfaction with the AI coach. Univariate outliers were identified as having z scores greater than 3. When outliers were present, univariate analyses were run without outliers. Multilevel modeling (MLM) was used to examine the main study hypotheses. MLM was used because (1) it allows the inclusion of all participants, regardless of missing data, while efficiently handling missing data; and (2) it is considered the preferred method to analyze longitudinal psychiatric data [30]. The linear mixed-effects models procedure of SPSS 28 was used for these analyses. For depression on the PHQ-8, a two-level MLM model using maximum likelihood estimation was computed. At Level 1, depression symptoms varied within participants over time as a function of a person-specific growth curve. This level tested for time effects, estimating the change in depression across the pre, postintervention, and follow-up assessment points. Time was coded such that the intercept reflected the initial level of depression at preintervention. Similar MLM models were computed for anxiety, hopelessness, self-efficacy, and agency. For each MLM model, we also computed the proportional reduction explained at the individual level, conceptualized as a change in pseudo R² when the time was added to the empty model (ie, random intercept-only model without any predictors). We used the following Snijders and Bosker [31] formula: R²=1−(σ²_F + τ²_F/ σ²_E + τ²_E), where σ²_F represents the level-one random error variance for the full model with time; τ²_F represents the level-two random error variance for the full model; σ²_E represents the level-one random error variance for the empty model; and τ²_E represents the level-two random error variance for the empty model. The effect size measure related to the variance explained for the overall model is ⨍²=R²/1−R² [32,33]. Guidelines for interpretation of ⨍² indicate that 0.02 is a small effect, 0.15 is a medium effect, and 0.35 is a large effect [34]. Finally, for outcome measures assessed at only two time points, such as well-being, 2-tailed paired samples t tests were used to examine within-participant change. Cohen d was calculated as a measure of effect size, with values of 0.2, 0.5, and 0.8 interpreted as small, medium, and large effects, respectively [32].

Of the 113 students who were eligible to participate, 50 students completed the pre- and postintervention surveys and at least 1 session with the app, and 41 students completed the 1-week follow-up survey. No significant differences were found between those who completed the follow-up survey and those who did not in terms of gender or baseline severity of depression or anxiety symptoms. A significant difference was found in terms of age (t₄₈= 2.42; P=.02), such that participants who did not complete the follow-up survey were significantly older (mean age 25, SD 6.41 y) than participants who completed the follow-up survey (mean age 21.4, SD 3.51 y). However, this difference is relatively modest and may not represent a clinically meaningful distinction among this age group. Accuracy checks assessing the range of summed scores (minimum and maximum scores) for each scale indicated that participants accurately completed the scales.

Of the 50 students who completed at least one session, 90% (n=45) were female. The mean age of the sample was 22.12 (SD 4.42) years. About 46% (n=23) self-identified as Latinx, 34% (n=17) non-Latinx White, 8% (n=4) non-Latinx Black, 8% (n=4) non-Latinx Asian, and 4% (n=2) multiracial. Most students were seniors (n=14, 28%). Half the students self-identified as heterosexual or straight (n=25). Most students were enrolled full-time (n=46, 92%), and 58% (n=29) were commuters or lived off-campus. Most students (n=36, 72%) also reported receiving federal or state financial assistance for their college education and being first-generation college students (n=29, 58%; see Table 1 for demographic characteristics of the sample).

The mean baseline PHQ-8 and GAD-7 scores were 13.50 (SD 5.06; range 4‐21) and 14 (SD 4.03; range 3‐23), respectively, with 80% (n=40) of students scoring above the clinical cutoff on the PHQ-8 (score ≥10) and 88% (n=44) on GAD-7 (score ≥10). The average hopelessness score was 5.38 (SD 2.67; range 0‐12). The mean agency score was 14.14 (SD 5.29; range 3‐23), self-efficacy was 15.98 (SD 3.98; range 4‐23), and wellness was 19.50 (SD 4.52; range 4‐29). Table 2 presents the mean score for each outcome of interest across the three study time points.

The mean number of sessions attended was 2.02 (SD 1.86; range 1‐10), suggesting that, on average, students completed around two sessions over the 1-week period with the AI mental health wellness coach. Of the four coach characters offered, the mean number of coaches used was 1.14 (SD 0.45; range 1‐3), which may suggest that students often opted to use the same coach character across sessions. Specifically, Michelle was selected a total of 19 times, Marcela was selected a total of 20 times, Professor Wilson was selected a total of 16 times, and Coach Thompson was selected a total of 2 times. On average, students used 1.1 (SD 1.13; range 0‐5) CBT-based skills during their conversations (ie, grounding techniques, mindfulness exercises, and exploring and challenging cognitive distortions). One univariate outlier was identified with regard to minutes spent per session (ie, z score above 3). After removing this participant from the analysis, the average minutes students spent in each session was 15.43 (SD 10.11; range 3.18‐53.48). Two outliers were also present with regard to message count. The average message count was 29.08 (SD 22.43; range 5‐115) after these two students were removed from the analysis. Finally, two outliers were identified for students’ response time. The students’ mean response time to each AI-generated response was 7.33 (SD 24.55; range 0.13‐143.37) minutes once these cases were excluded.

A linear model with random slopes and intercepts revealed that students’ GAD-7 scores decreased significantly from preintervention to 1-week follow-up (β=−2.15; P<.001). About 14% of the variance in anxiety was explained by time (R²=0.13; ⨍²=0.14). Similarly, there was a significant reduction in students’ PHQ-8 (β=−1.62; P<.001) and hopelessness scores (β=−.64; P<.001), with time accounting for 8% and 3% of the variance, respectively (R²=0.07, ⨍²=0.08; R²=0.02, ⨍²=0.03, respectively). With regard to agency and efficacy, students’ scores significantly increased from preintervention to follow-up (β=.64, P=.32; β=.53, P=.02, respectively). However, time accounted for less than 1% of the variance in agency (R²=0.007; ⨍²=0.007) and efficacy (R²=0.009 ⨍²=0.009). See Table 3 and Multimedia Appendix 1 for results of MLM analyses. To examine changes in well-being, a 2-tailed paired samples t test was computed. Students’ well-being scores significantly increased from preintervention (mean 18.98, SD 4.35) to follow-up (mean 21.66, SD 0.79; t₄₀=2.90; P=.006, d=0.45). Relatedly, at postintervention, 62% (n=31) of students reported that their mental wellness was better or much better after one session with Wayhaven. As a sensitivity analysis, the current mental health treatment status was added as a covariate to the linear mixed models. This variable did not meaningfully affect the pattern or significance of the results. Thus, it was not included in the final models reported here.

Immediately after using Wayhaven (ie, at postintervention), participants generally endorsed that they found Wayhaven to be acceptable. Specifically, 90% (n=45) of participants agreed that Wayhaven was easy to use, and 74% (n=37) of participants indicated that they were satisfied with their experience with Wayhaven. Most participants (n=36, 72%) also agreed that they would use Wayhaven again in the future. The majority of students (n=42, 84%) participants also agreed that there is a need for a resource like Wayhaven among college students. Many students also reported finding the app quite engaging (n=26, 53%).

At 1-week follow-up, there was a slight decrease in students’ satisfaction, with 65% (27/41) of participants indicating that they were satisfied with their experience using Wayhaven. Additionally, 59% (24/41) of participants indicated that they would use Wayhaven again in the future. Despite these shifts, 68% (27/41) of participants noted that they would recommend Wayhaven to a friend who may be experiencing similar mental wellness concerns. Participants also endorsed the strong feasibility of Wayhaven, such that 80% (33/41) of participants felt that Wayhaven fit easily into their daily routine (see Table 4 for detailed frequency of student responses to satisfaction questions at postintervention and follow-up). More than half of the students also continued to find the app quite engaging (22/41, 56%).

The purpose of this study was to explore how college students interact with Wayhaven’s AI-powered mental wellness chatbot and its feasibility for improving mental wellness among college students with elevated depression or anxiety symptoms. Specifically, this study sought to assess the feasibility and acceptability (ie, engagement and satisfaction) of using Wayhaven with students with elevated depression or anxiety symptoms, as well as whether app use was linked with improvement across several mental health outcomes over a 1-week period. Most participants self-identified as racially and ethnically minoritized and were experiencing clinically elevated depression or anxiety symptoms. Overall, Wayhaven use was associated with a significant decrease in depression, anxiety, and hopelessness from preintervention to follow-up. Wayhaven use was also related to a significant increase in agency, efficacy, and well-being across the study period. Additionally, students reported being satisfied with Wayhaven and seeing the app as a new important resource for college students.

More than half of the participants scored above the clinical cutoff on the PHQ-8 and GAD-7, which is in line with previous research documenting elevated rates of internalizing symptoms among college students [2]. It is clear that there is a need for increased access to efficacious and equitable mental health resources for college students, particularly those self-identifying as racially and ethnically minoritized [35,36]. Of note, 66% (n=33) of students in our sample were racially and ethnically minoritized. DMHIs may be particularly appealing to racially and ethnically minoritized students who are more likely than White students to indicate a preference for web-based treatment versus in-person therapy [37,38]. Our findings suggest that Wayhaven may be a viable mental health resource to such populations, given the reduction in depression, anxiety, and hopelessness scores in a relatively short period of time.

Results also suggest that the majority of students were satisfied with their experience using Wayhaven. Most students felt that there was a need for a resource like Wayhaven among college students and that they would recommend it to their peers. With regard to the coach personalities offered, most students opted to use the same coach persona across sessions and selected Marcela, the Latinx college student, most frequently. This preference among students aligns with previous research suggesting that college students tend to favor seeking help from informal sources, like peers [39]. In addition to rating the app as engaging, students completed two sessions on the app on average across the 1-week study period, with the average message count across sessions being around 29 (SD 22.43). The number of messages students exchanged is comparable to those reported by He et al [40], who tested a cognitive behavioral therapy−based mental health chatbot, XiaoE, for young adults with depressive symptoms. They found that the average number of messages exchanged with XiaoE was 25.54 (SD 26.25) daily across a 1-week period. It should be noted that participants in this study were instructed and prompted to engage with XiaoE daily, whereas students in this study were asked to use Wayhaven as needed. It is unclear whether user engagement (ie, sessions and messages) would be higher if students were asked to use the tool daily. Taken together, these data suggest that Wayhaven appears to be engaging, acceptable, and feasible for college students with elevated depression or anxiety symptoms.

While these preliminary findings support the applicability of Wayhaven in college populations, this study is not without its limitations; caution should be used when considering the generalizability of the findings. This was a small, uncontrolled pilot study, and we did not include a control group. Therefore, we cannot infer that the observed symptom improvements were a result of the intervention. However, our goal was not to conduct a randomized controlled trial (RCT); instead, our aim was to test the feasibility and acceptability of a novel AI mental wellness chatbot in a real-world setting. Testing an intervention’s feasibility and acceptability via an open trial is an important first step prior to an RCT [41]. As noted above, this research should be followed up with a larger RCT comparing Wayhaven to an active control condition matched by age and sex and a longer follow-up period. The use of self-report assessments is also limited in that participants may have under- or overreported the severity of their symptoms. Importantly, self-report questionnaires of depression and anxiety symptoms have been found to be moderately to strongly correlated with clinician-rated scales, and previous work using these measures has proven to be clinically impactful [42,43]. The majority of participants were also female, possibly limiting generalizability to male participants. Additionally, our sample was racially and ethnically diverse, participants were obtained from a single public university in New Jersey, and results may not be generalizable to other areas of the United States.

Although Wayhaven use was associated with a statistically significant improvement in outcome scores across time, the magnitude of these effects was small. The observed effect sizes were strongest for anxiety, followed by depression, which were our main clinical outcomes. These findings are consistent with prior research on brief interventions that suggests such programs have the strongest effects on anxiety symptoms [44,45]. Researchers have noted that categorizing effect sizes as “small,” “medium,” or “large” is not inherently meaningful without a contextual frame of reference [46]. In public health, even small effects can yield a substantial impact when scaled across populations. Small shifts in mean scores can disproportionately affect the extreme end of a population distribution and drive higher rates of clinical referrals, as illustrated by the surge in mental health service use among youth during the COVID-19 pandemic despite only a modest increase in mental health outcome scores [47]. Moreover, brief interventions with small average effects may still yield substantial population-level benefits by offering a low-cost means to expand the reach and impact of mental health services, particularly for individuals in need, like college students who may be reluctant to seek higher-intensity support [45,48]. Nevertheless, future research should explore ways to strengthen and sustain the impact of Wayhaven, such as incorporating booster sessions or embedding the intervention within a broader stepped-care framework.

Our findings mirror previous research on the effectiveness of AI chatbots in reducing depression and anxiety symptoms among college students [13,14,41]. However, in contrast to previous studies testing scripted chatbots, Wayhaven’s generative AI-powered mental wellness chatbot offered students personalized, context-sensitive support by integrating demographic information, university-specific resources, and a customized AI coach persona. Given our preliminary findings, future research should focus on comparing the effectiveness of Wayhaven to an active control condition, such as a mental health app or internet-based intervention, and among a larger sample of students. If proven effective, Wayhaven can serve as a lower-intensity intervention that college counseling centers can offer to students presenting with subclinical or acute depression or anxiety symptoms, which can help free in-person services to students experiencing more severe symptoms or distress.

Despite limitations, results revealed that Wayhaven led to an improvement in student mental well-being outcomes (ie, depression, anxiety, hopelessness, agency, efficacy, and well-being) and students were satisfied with the tool. Wayhaven may be a promising resource to deliver brief, evidence-based text conversations to support college students’ mental wellness needs. Future research should examine the effectiveness of Wayhaven by comparing it to an active control condition and including a larger student sample. With the already overwhelming and unmet mental health needs of this population, a DMHI like Wayhaven may offer a pathway to free, accessible, and equitable mental wellness care.