Help-seeking attitudes, stigma and artificial intelligence counseling preferences among university students

Abstract

BACKGROUND

High rates of psychological distress among university students, together with persistently low use of professional mental health services, have become a notable public health issue.

AIM

To examine how help-seeking attitudes, stigma and individual psychological characteristics jointly shape university students’ preferences for different counseling formats.

METHODS

A cross-sectional survey was conducted which included 313 university students. Latent profile analysis was applied to identify distinct subgroups based on help-seeking attitudes and perceived stigma. Differences in psychological symptoms and preferences for artificial intelligence (AI) counseling were then compared across these subgroups.

RESULTS

The following three subgroups were identified: “Low willingness-high stigma” (36.4%), “moderate willingness-moderate stigma” (43.8%), and “high willingness-low stigma” (19.8%). Interestingly, the “low willingness-high stigma” group showed the most severe depression and anxiety symptoms, and reported the highest prior use of AI-based support and the strongest preference for AI counseling.

CONCLUSION

AI and human counselors play complementary roles: AI offers advantages in accessibility and anonymity, while human counselors remain irreplaceable in terms of empathy, complex problem-solving, and long-term therapeutic relationships.

Key Words: Artificial intelligence; Psychological counseling; Help-seeking attitudes; Stigma; Latent profile analysis

Core Tip: Students who report the most severe symptoms and the highest levels of stigma are actually the ones who show the strongest preference for artificial intelligence (AI)-based counseling. This suggests AI can serve as a low-stigma, low-barrier entry point for those reluctant to seek traditional help. Meanwhile, human counselors still play an irreplaceable role in deeper therapeutic work. An integrated model combining both approaches is needed.

INTRODUCTION

The mental health of university students has emerged as a pressing global public health challenge, with a high prevalence of psychological symptoms documented worldwide. An umbrella review of 62 meta-analyses, encompassing over 8.7 million university students, reported prevalence rates of 35.41% for mild depressive symptoms, 40.21% for mild anxiety symptoms and 36.34% for stress symptoms, with moderate-to-severe depression and anxiety affecting 13.42% and 16.79% of this population, respectively[1]. Against this backdrop of substantial mental health need, a recent systematic review and meta-analysis found that only 28% of students experiencing psychological distress actively seek help, a pronounced gap between help-seeking intentions and actual behavior[2]. Commonly cited barriers include a preference for self-management, time constraints, limited awareness of available resources, and the perception that professional help is unnecessary[2].

Concurrent with these unmet needs, advances in artificial intelligence (AI) have positioned AI-based tools as an emerging complement to traditional mental health support for young people. In a survey which included 500 individuals aged 16 years to 90 years, 74.6% reported using ChatGPT, and of those, 17.2% had turned to it for mental health concerns, a proportion that climbed to 39.5% among adolescents, markedly higher than the rate seen in older age groups[3]. For younger users, the main reasons for seeking AI support were to gain coping strategies (73.3%), explore different perspectives (61.6%) and organize their thoughts (53.5%); over half (55.8%) rated AI’s help as moderately to highly useful[3]. These findings point to the increasing significance of AI-based chatbots as an emotional support resource for adolescents and young adults.

Attitudes toward professional psychological help reflect a person’s tendency to seek such support when their own coping resources fall short. Studies consistently find that more favorable help-seeking attitudes are linked to actual help-seeking behavior[4] and a greater willingness to self-disclose[5] and they tend to support better psychological outcomes[6]. Corrigan[7] described stigma as an integrated construct comprising public stigma, negative stereotypes held by society toward stigmatized groups, and self-stigma, which arises when individuals internalize those views, leading to self-devaluation and a reduced sense of efficacy. In research on stigma and psychological help-seeking, two commonly used measures are the Stigma Scale for Receiving Psychological Help (SSRPH), developed by Komiya et al[8], and the Self-Stigma of Seeking Help Scale (SSOSH), developed by Vogel et al[9]. A number of studies indicate that higher self-stigma is associated with lower self-appraisal and less positive help-seeking attitudes, and it acts as a notable obstacle to using mental health services[10-12].

Thus, both stigma and attitudes toward help-seeking are important determinants of how individuals choose between different mental health support options[13]. As AI-based and traditional face-to-face counseling increasingly coexist, university students’ preferences in this area show considerable complexity and variability. Media platforms may help reduce risk and can play a meaningful role in shaping health-related behaviors and outcomes[14]. Previous work also suggests that human-computer interaction is a key factor in user acceptance[15], and that psychological safety and trust are central to the human-AI counseling relationship[16]. One study found that students with higher mental health literacy were more likely to seek professional help during the pandemic[17], suggesting that help-seeking attitudes and behaviors are shaped by multiple interacting factors. Despite growing attention on digital mental health, little is known about the psychological processes underlying university students’ choices between AI-based and human counselors particularly in the Chinese context. To address this gap, the current study examines how help-seeking attitudes, stigma and individual psychological characteristics jointly shape university students’ preferences for different counseling formats. The findings are expected to offer theoretical insight and practical guidance for the design of more targeted mental health services in higher education settings.

MATERIALS AND METHODS

Participants

Participants were recruited from among full-time undergraduate students between October and December 2025 via poster advertisements placed at three universities in Shanghai. Interested students self-enrolled by scanning a QR code that directed them to an online informed consent form and a subsequent questionnaire hosted on the Wenjuanxing platform. Eligibility criteria included: (1) Being a full-time undergraduate student; (2) Aged 18 years or older; and (3) Providing valid electronic informed consent. All questionnaire items were set as required fields. By December 31, 2025, a total of 313 students completed the questionnaire and were retained in the final analytical dataset. The final sample comprised 160 female (51.1%) and 153 male (48.9%) participants. The average age of the students was 20.9 years. The research protocol was reviewed and approved by the Institutional Review Board of Naval Medical University in China. Electronic informed consent was obtained from all individuals prior to data collection.

Measures

Generalized Anxiety Disorder Scale: Based on the diagnostic criteria for generalized anxiety disorder from the Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV, the Generalized Anxiety Disorder Scale-7 (GAD-7) consists of seven items, each rated on a 4-point scale ranging from 0 (not at all) to 3 (nearly every day)[18]. Total scores range from 0 to 21, with the following standard cut-off scores: 0-4, no anxiety; 5-9, mild anxiety; 10-14, moderate anxiety; ≥ 15, severe anxiety. The GAD-7 has demonstrated strong psychometric properties across diverse populations, including satisfactory internal consistency and good diagnostic accuracy for identifying anxiety disorders[19].

Patient Health Questionnaire-9: The Patient Health Questionnaire-9 (PHQ-9) is one of the most widely used, reliable and validated measures of depressive symptoms[20]. This nine-item self-report questionnaire assesses the frequency and severity of depressive symptoms over the past two weeks according to the criteria for major depressive disorder in the DSM-IV. Items are rated on a 4-point scale ranging from 0 (not at all) to 3 (nearly every day). Total scores between 0 and 4 indicate no depressive symptoms. In the present study, the Cronbach’s α coefficient for this scale was 0.93.

Attitudes Toward Seeking Professional Psychological Help Scale: The Attitudes Toward Seeking Professional Psychological Help Scale (ATSPPH) was originally developed by Fischer and Turner[21] to assess individuals’ general attitudes toward seeking professional psychological assistance. The short-form Chinese version was translated and adapted by Hao and Liang[22] for use in Chinese populations. The scale comprises ten items rated on a 4-point Likert scale ranging from 0 (strongly disagree) to 3 (strongly agree). Total scores range from 0 to 30, with higher scores indicating more positive help-seeking attitudes (0-15, negative; 16-22, moderate; 23-30, positive). The instrument captures two underlying dimensions: Acceptance of seeking treatment for emotional problems and perceived value or need for psychological support. In the present study, the overall scale demonstrated acceptable internal consistency (Cronbach’s α = 0.784), with subscale coefficients of 0.821 for acceptance of seeking treatment and 0.815 for values or need for psychological support.

SSRPH: The SSRPH was developed by Komiya et al[8] to assess individuals’ perceived public stigma associated with seeking professional psychological help, defined as negative stereotypes held by the broader society toward stigmatised groups who seek psychological assistance. The scale consists of five items rated on a 5-point Likert scale ranging from 0 (strongly disagree) to 4 (strongly agree), with total scores ranging from 0 to 20. Higher scores indicate higher levels of perceived public stigma. This scale has been widely used in cross-cultural research on stigma related to psychological help-seeking, and the Chinese version has demonstrated sound psychometric properties among Chinese college students[23]. In the present study, the Cronbach’s α coefficient for this scale was 0.942.

SSOSH: The SSOSH was developed by Vogel et al[9] to measure the extent to which individuals internalise public stigma into self-stigma, manifested as self-devaluation and self-inefficacy when seeking psychological help. The scale consists of ten items rated on a 5-point Likert scale ranging from 0 (strongly disagree) to 4 (strongly agree), with total scores ranging from 0 to 40. Higher scores indicate higher levels of self-stigma. In the present study, the Cronbach’s α coefficient for this scale was 0.764.

AI related variables: All AI relevant items were developed for the purposes of this study based on a review of the emerging literature on digital mental health and human-AI interaction. The following single item measures were used: (1) Prior use of AI-based psychological support, assessed by the question “Have you ever used any AI-based tool or application for psychological support or emotional relief?” with a binary yes/no response; (2) The specific AI tools that participants had previously used, collected via an open ended text entry question; (3) Future counseling preference, measured by a single question with five ordered response options ranging from “rely solely on AI tools” to “rely exclusively on human professional counselors”, with three intermediate mixed use options (AI first, equal preference, and human first); and (4) Comparative ratings of AI vs human counselors across seven domains, each rated on a 5-point Likert scale ranging from 1 (“human counselor is much better”) to 5 (“AI is much better”). No composite scores were computed; all analyses used the original item level responses.

Statistical analysis

To identify latent classes of help-seeking attitudes and stigma among university students, we conducted latent profile analysis. To identify latent subgroups based on attitudes toward seeking professional psychological help and stigma, a series of latent profile analyses were conducted, estimating models with one to four classes. The indicators entered into the model included scores on each subdimension of the ATSPPH scale (acceptance of seeking treatment and perceived need for psychological support) and scores on the stigma measure (self-stigma and public stigma). All continuous indicators were standardized (z-scores) prior to analysis to ensure comparability across different metrics. Model selection was based on a combination of fit indices including the Akaike information criterion (AIC), Bayesian information criterion (BIC), sample-size-adjusted BIC (aBIC), entropy, the Lo-Mendell-Rubin adjusted likelihood ratio test (LMR), and the Bootstrap likelihood ratio test along with considerations of parsimony and interpretability. When the optimal class solution was determined, we used one-way analysis of variance and χ² tests to compare classes on mental health status, preferences for psychological help-seeking, and future help-seeking intentions. Post hoc pairwise comparisons were adjusted with the Bonferroni correction. All analyses were performed using Mplus 8.0 and SPSS 26.0, with a two-tailed significance threshold of P < 0.05.

RESULTS

Participant characteristics and AI-based mental health support use

A total of 313 university students participated in this study, with a mean age of 20.90 years (SD = 2.28). The sample comprised 153 male (48.9%) and 160 female (51.1%) participants. Non-only-child status was reported by 187 participants (59.7%). The majority of participants resided in urban areas (n = 226, 72.2%) and self-rated their family economic status as moderate (n = 204, 65.2%). Among the 313 participants, 95.5% reported having used AI tools for general purposes in their daily lives. However, when specifically asked about AI-based psychological support, only 32.3% (approximately one-third) indicated that they had used such tools for mental health or emotional relief. Among those who had used AI for psychological support, the most frequently reported tools included DeepSeek (49.5%) and Doubao (42.6%). In terms of usage patterns, nearly half of the users (49.5%) had used AI tools only once or twice in total, while 29.7% reported turning to such tools solely during episodes of acute psychological distress. In terms of session duration, the vast majority (85.1%) engaged in interactions lasting 15 minutes or less per session (Table 1).

Table 1 Demographic characteristics of participants and their use of artificial intelligence psychological support tools, mean ± SD/n (%).

Characteristics
Age	20.9 ± 2.28
Gender
Male	153 (48.9)
Female	160 (51.1)
Only child status
Only child	126 (40.3)
Not only child	187 (59.7)
Family location
Town	226 (72.2)
Rural	87 (27.8)
Family economic status
Poor	29 (9.3)
Average	204 (65.2)
Good	80 (25.5)
AI tool usage
Ever used AI tools (general purposes)	299 (95.5)
Ever used AI tools for psychological support	101 (32.3)
Primary AI psychological help tool
DeepSeek	50 (49.5)
Doubao	43 (42.6)
Others	8 (7.9)
Usage frequency
Daily	6 (5.9)
Weekly	12 (11.9)
Monthly	3 (3.0)
As needed	30 (29.7)
Once/twice	50 (49.5)
Duration per session
< 15 minutes	86 (85.1)
15-30 minutes	10 (9.9)
> 30 minutes	5 (5.0)

AI: Artificial intelligence.

Perceived efficacy of AI-based psychological support and comparison with human counselors

Independent-sample t tests were conducted to compare prior users of AI-based psychological support (n = 101) and non-users (n = 212) on their relative evaluations of AI vs human counselors across multiple dimensions. Significant differences in perceived ratings were observed between the two groups on several core dimensions (Figure 1). On the dimension of listening patience, users rated AI significantly higher (M = 3.37, SD = 1.34) than non-users (M = 2.87, SD = 1.23). Notably, users perceived AI as superior in this aspect, whereas non-users reported no significant difference between AI and human counselors. Regarding depth of empathy and understanding, although both groups considered human counselors to be more advantageous, users’ ratings (M = 2.34, SD = 1.31) were significantly higher than those of non-users (M = 1.85, SD = 0.99), P < 0.001. For creating a safe and non-judgmental environment, users also reported significantly higher ratings [M = 3.28, SD = 1.30, 95% confidence interval (CI): 3.03-3.53] compared to non-users (M = 2.82, SD = 1.16, 95%CI: 2.66-2.98), with the two groups diverging in their attribution of superiority: Users perceived AI as superior, whereas non-users tended to favor human counselors. No statistically significant differences were found between the two groups on the dimensions of professional insights and solutions, ability to handle complex issues, or long-term therapeutic reliability (all P > 0.05). Across all participants, professional insights and solutions, and long-term therapeutic reliability were consistently identified as areas of clear human advantage. On the dimension of accessibility and convenience, participants across both groups unanimously recognized this as an area of clear AI advantage.

Figure 1 Mean comparative ratings of artificial intelligence and human counselors among prior users and non-users of artificial intelligence-based psychological support. ^bP < 0.01. NS: Non significance.

Latent profile analysis of help-seeking attitudes and stigma and its association with AI-based mental health support preferences

To identify latent subgroups based on attitudes toward seeking professional psychological help and stigma, a series of latent profile analyses were conducted, estimating models with one to four classes. Fit indices for each competing model are presented in Table 2. Overall, the information criteria including AIC, BIC, and aBIC decreased progressively as the number of classes increased, indicating improved model fit with greater complexity. Although the four-class model yielded a slightly higher entropy value, the LMR test for the four-class solution was not statistically significant (P = 0.086), suggesting that adding a fourth class did not significantly improve model fit over the three-class solution. In addition, the four-class model contained a class with a very small membership probability (6%), raising concerns about overfitting and practical interpretability. In contrast, the three-class model demonstrated a statistically significant LMR test (P = 0.004) and a well-balanced class distribution (36%, 44%, and 20%), with no excessively small classes (Figure 2). The Bootstrap Likelihood Ratio Test was also significant for the three-class model (P < 0.001), further supporting its superiority over the two-class solution. Considering these statistical criteria together with the principle of parsimony, we retained the three class latent profile model as the final solution. Subsequent analyses were conducted to characterize and compare the identified subgroups.

Figure 2  Indicator profiles across the three latent classes.

Table 2 Fit indices for latent profile models of help-seeking attitudes and stigma.

Model	AIC	BIC	aBIC	Entropy	BLRT P value	LMR P value	Class probability (%)
1	6828.86	6858.83	6833.46				1
2	6608.45	6657.15	6615.91	0.73	< 0.001	< 0.001	0.53/0.47
3	6493.93	6561.36	6504.27	0.88	< 0.001	0.004	0.20/0.36/0.44
4	6410.99	6497.15	6424.20	0.90	< 0.001	0.086	0.32/0.42/0.06/0.20

AIC: Akaike information criterion; BIC: Bayesian information criterion; aBIC: Adjusted Bayesian information criterion; BLRT: Bootstrap Likelihood Ratio Test; LMR: Lo-Mendell-Rubin adjusted likelihood ratio test.

To examine differences across the three latent classes on key variables, we compared their characteristics in terms of mental health status and preferences for AI-based psychological counseling (Table 3). Based on levels of help-seeking willingness and stigma, the three classes were labeled as the “low willingness-high stigma” group (n = 114), the “moderate willingness-moderate stigma” group (n = 137), and the “high willingness-low stigma” group (n = 62). Significant group differences emerged in mental health status. The “low willingness-high stigma” group scored significantly higher on the PHQ-9 than the other two groups (F = 14.502, P < 0.001). This group also showed a significantly higher proportion of individuals with mild or greater depressive symptoms (60.5%) than the moderate willingness-moderate stigma group (35.8%) and the high willingness-low stigma group (29.0%; χ² = 20.066, P < 0.001). A similar pattern held for GAD-7 total scores and the proportion with mild or greater anxiety symptoms (total score: F = 13.129, P < 0.001; proportion: χ² = 25.632, P < 0.001), with the low willingness-high stigma group again exhibiting the most severe symptom levels.

Table 3 Mental health status and artificial intelligence help-seeking preferences by latent profile (based on help-seeking attitudes and stigma), mean ± SD/n (%).

	Low willingness and high stigma (n = 114)	Moderate willingness and moderate stigma (n = 137)	High willingness and low stigma (n = 62)	F/χ2 value	P value
PHQ-9 total score	6.55 ± 6.16	3.94 ± 4.44	2.71 ± 3.31	14.502	< 0.001
0-4	45 (39.5)	88 (64.2)	44 (71.0)	20.066	< 0.001
≥ 5	69 (60.5)	49 (35.8)	18 (29.0)
GAD-7 total score	5.21 ± 5.26	2.87 ± 4.04	2.13 ± 3.10	13.129	< 0.001
0-4	55 (48.2)	105 (76.6)	47 (75.8)	25.632	< 0.001
≥ 5	59 (51.8)	32 (23.4)	15 (24.2)
Ever used AI tools for psychological support
No	67 (58.8)	102 (74.5)	43 (69.4)	7.093	0.029
Yes	47 (41.2)	35 (25.5)	19 (30.6)
Future help-seeking preference
AI-preferred	29 (25.4)	19 (13.9)	3 (4.8)	14.997	0.005
Balanced	24 (21.1)	26 (19.0)	17 (27.4)
Human-preferred	61 (53.5)	92 (67.1)	42 (67.8)

Differences across groups also emerged in the use of and preferences for AI-based psychological support. Prior experience with AI-based support was more common in the low willingness-high stigma group (41.2%) than in the moderate willingness-moderate stigma group (25.5%; χ² = 7.093, P = 0.029). When asked about future counseling preferences, the low willingness-high stigma group expressed a stronger inclination toward AI-based counseling (25.4%) than the other two groups (13.9% and 4.8%; χ² = 14.997, P = 0.005). In contrast, the moderate willingness-moderate stigma and high willingness-low stigma groups both favored human-delivered counseling to a greater extent.

DISCUSSION

University students’ engagement with AI-based psychological counseling shows a clear pattern: Acceptance is high, but actual use is infrequent and highly dependent on context. Help-seeking willingness also varies considerably across subgroups. While over 95.5% of students reported using AI tools, only about a third did so for mental health purposes, and most interactions were brief. This aligns with Pandi-Perumal et al[24], who noted that AI applications in mental health are still moving from development to practical implementation.

AI received higher ratings than human counselors on accessibility, convenience, and anonymity. By contrast, human counselors were seen as irreplaceable when it comes to empathic depth, professional problem-solving, and building long-term therapeutic relationships. This perception of “functional complementarity” rather than “simple substitution” echoes the perspective of Gültekin and Şahin[25], who conceptualized AI as serving an “assistant” rather than a “replacement” role in mental health care. AI’s strengths were clearly evident in situations where immediacy, anonymity, or reduced stigma mattered: Late-night conversations or concerns involving sensitive privacy, for instance. In such cases, AI-based support may fill gaps that traditional services are less able to address.

The three subgroups identified by latent profile analysis were: High willingness-low stigma, high stigma-low willingness, and moderate willingness-moderate stigma. Notably, the high stigma-low willingness group reported the most severe psychological symptoms yet showed the strongest preference for AI as a primary support channel. At first glance, this seems to run counter to Clement et al’s argument[26] that stigma discourages professional help-seeking. However, rather than contradict it, the finding points to a shift in how stigma operates in the digital age.

Our findings tentatively suggest that, among students who are deterred from seeking traditional help because of shame or stigma, AI based tools might be perceived as a more accessible or less intimidating alternative even for those who hold reservations about the technology itself. In our sample, a considerable number of participants reported using digital tools to cope with psychological distress, which appears consistent with the broader observation that such platforms are becoming a common resource in university settings. We also note that the current widespread availability of AI platforms may be associated with help seeking choices among stigmatized students, although this possibility should be tested in future longitudinal or experimental studies. Theoretically, these subgroup-specific findings contribute to a more nuanced understanding of help-seeking heterogeneity and suggest that future personalized mental health services might benefit from subgroup-specific targeting rather than a one-size-fits-all approach. Empirically, they provide preliminary evidence that can inform the differentiated design of AI-based mental health interventions. These interpretations are put forward as hypotheses generated from our data, rather than as conclusive statements, and should be interpreted with caution given the study design.

These findings point to both the limitations and the potential of AI in mental health care[27]. Algorithmic bias, questions about long-term effectiveness, and unresolved ethical issues remain real barriers to advancing AI-based psychological services[28]. Moving forward, two priorities stand out. One is improving AI’s capacity for emotional interaction and contextual understanding, especially when it comes to empathy and dealing with complex psychological concerns[29]. The other is building a stratified, integrated service system that lets AI and human counselors work alongside each other. In such a setup, AI could handle initial screening, immediate response, and routine psychoeducation, leaving in-depth intervention, crisis management, and longer-term therapy to human professionals. Encouraging more positive attitudes toward AI-based counseling and steadily improving service quality are also important. For groups with high stigma, emphasizing privacy protection and other strengths of AI-based services may help reduce stigma levels[30].

This study has some limitations. First, the cross-sectional design makes it difficult to determine causal relationships; experimental or longitudinal studies are needed to clarify the underlying mechanisms. Second, the AI platforms used by participants were general-purpose chatbots, not specialized mental health AI applications. Future studies should examine whether dedicated mental health AI tools produce different patterns of help-seeking preferences. Third, the ethical and practical risks associated with AI-based mental health support, including privacy and data security concerns, the lack of crisis detection mechanisms, and the potential for delayed professional help-seeking were not directly examined in this study and warrant careful investigation in future research. Fourth, the primary outcome measure of future counseling preference was a single, non-validated item. Future research should develop and validate multi-dimensional instruments to assess preferences for AI vs human counseling. Finally, as cultural factors and religious beliefs may influence help-seeking preferences[31], cross-cultural comparisons will be necessary to determine how well these findings generalize. Future studies might examine how personality traits, types of psychological problems, or specific counseling topics shape preferences for different modalities. There is also room for empirical research comparing the effectiveness of AI-based vs human-delivered counseling, particularly in terms of symptom relief and changes in help-seeking behavior. Such work would be a valuable extension of the current study.

CONCLUSION

This study investigated how attitudes toward professional help-seeking and stigma relate to preferences for AI-based psychological support among 313 university students. Through latent profile analysis, we identified three subgroups: A high-willingness/low-stigma group (19.8%), a moderate group (43.8%), and a low-willingness/high-stigma group (36.4%). Interestingly, the preference for AI-oriented support was significantly pronounced only in the low-willingness/high-stigma subgroup, implying that AI tools may be particularly appealing to individuals who encounter multiple obstacles in accessing conventional mental health services. These results suggest that university mental health services should acknowledge the diversity in students’ help-seeking profiles and consider AI tools as complementary bridges to professional care while also remaining cautious about the risk that heavy reliance on AI could postpone necessary clinical intervention. To build on these findings, future work would benefit from longitudinal designs and the development of validated measures tailored to this emerging area.

ACKNOWLEDGEMENTS

The authors would like to acknowledge the volunteers who participated in the study.