In recent years, the mental health of university students has become a growing concern worldwide (Ferrari et al., 2022). A report from the World Health Organization (WHO) indicates that depression has become one of the leading causes of disability and death among adolescents. In China, the Ministry of Education and several epidemiological surveys show a year-on-year increase in the detection rate of depression among university students, with some regions exceeding 20% (Lin and Li, 2025). However, faced with this massive demand, the traditional mental health service system is confronted with severe challenges. On one hand, there is a serious shortage of psychological counselors in universities, making it difficult to cover all students in need (Juniar et al., 2025). On the other hand, constrained by traditional cultural norms and stigma, many students suffering from depressive symptoms would rather endure their pain alone than step into an offline counseling office (Li et al., 2024). This situation, characterized by both a supply–demand imbalance and barriers to seeking help, urgently calls for the emergence of new intervention models.

With breakthroughs in Generative AI technology, AI-empowered psychological therapy tools have emerged (Fitzpatrick et al., 2017; Ni and Jia, 2025). Among them, AI music therapy has attracted considerable attention due to its unique advantages. A large body of research has already confirmed the effectiveness of traditional music therapy in alleviating stress and depressive symptoms (De Witte et al., 2022; Liu et al., 2025). Unlike the passive listening of traditional music, modern AI systems (such as Suno, Mubert, etc.) can generate personalized therapeutic music in real-time based on the user’s emotional state, heart rate variability (HRV), or even brainwave data (Liu et al., 2024). For university students with depression, AI music therapy holds a natural appeal: first, its low barrier to entry and high accessibility allow students to access services anytime and anywhere with just a smartphone, breaking through the time and space constraints (Gbollie et al., 2023); second, the sense of security from depersonalized interaction helps avoid the evaluation anxiety present in interpersonal communication, providing a safe “psychological buffer zone” for depressed individuals with social withdrawal tendencies (Crawford et al., 2024). Therefore, AI music therapy is considered a potentially effective path to mitigating the campus mental health crisis (He et al., 2024). Specifically for depressed students, who often experience anhedonia (the inability to feel pleasure) and motivational deficits, the sensory and non-verbal nature of music offers a unique therapeutic modality that is less cognitively demanding than traditional talk-based therapies. It provides an alternative channel for emotional expression and regulation, making it a theoretically relevant intervention for this population.

However, technological development also brings corresponding challenges and risks. The high degree of personalization in AI music therapy relies on the deep mining of users’ sensitive data. To generate precise musical prescriptions, algorithms often need to collect highly private information such as users’ emotional records, physiological indicators, and even voice intonation. This collection of affective and physiological data is qualitatively different and potentially more intrusive than the text-based disclosures common in AI chatbots, raising unique privacy concerns. In the context of price discrimination based on big data analysis and frequent data breaches, this “panoptic surveillance” of one’s inner mental world has triggered widespread privacy anxiety (Barth and de Jong, 2017). For university students with depressive tendencies, who are already sensitive and suspicious, will this privacy concern become a key hindering factor? Specifically, because the data captured by AI music therapy (e.g., emotional, vocal, physiological) can be perceived as more revealing of one’s inner state than text-based data, understanding the role of privacy concerns becomes particularly critical. This is a pressing real-world question that needs to be answered (Li et al., 2024).

This study constructs an integrated analytical framework by introducing a psychopathological variable (depression) into the classic Technology Acceptance Model (TAM) and combining it with Privacy Calculus Theory. This not only extends the applicability of TAM to special populations (psychologically sub-healthy individuals) and specific contexts (digital healthcare) (Zhang et al., 2025) but also enriches our understanding of the “privacy paradox” in the field of health behavior (Barth and de Jong, 2017). More importantly, by identifying a key psychological boundary condition—privacy concern—this study specifies when psychological distress is more or less likely to translate into the intention to seek digital help. This moves beyond simply extending existing models and offers a more nuanced explanation for the complex decision-making process of vulnerable users, highlighting a critical point of friction in the adoption of digital mental health tools.

The findings of this study will directly serve the digital transformation of mental health services in universities. If privacy concerns are confirmed to be a key inhibiting factor, then universities and developers, when promoting AI tools, cannot simply focus on publicizing “powerful features” but must front-load “data security commitments.” The conclusions will provide strategic guidance not only for designing AI products that are more humanistic, but also for university health administrators to formulate transparent data governance policies, ensuring that digital interventions are both accessible and trustworthy.

The Technology Acceptance Model, first proposed by Davis (1989), is one of the most influential theoretical models for explaining users’ behavioral intentions to adopt information systems. Rooted in the Theory of Reasoned Action (TRA), TAM’s core proposition is that an individual’s Behavioral Intention to use a technology is primarily determined by two core beliefs:

Perceived Usefulness (PU): The degree to which a user believes that using a particular system will enhance their job performance or quality of life.

Perceived Ease of Use (PEOU): The degree to which a user believes that using the system will be free of effort.

In the mental health domain, TAM has been widely applied to explain the adoption of online counseling, mental health apps, and wearable devices (Gbollie et al., 2023; Kleine et al., 2025). However, the traditional TAM model assumes users are perfectly rational decision-makers, often overlooking the distorting effect of emotional states (such as anxiety, depression) on cognitive evaluations, and it does not fully consider external social environmental factors.

To address the shortcomings of the TAM model, Venkatesh et al. (2003) proposed the Unified Theory of Acceptance and Use of Technology (UTAUT). UTAUT integrates eight major models, including TAM, and proposes four core determinants: Performance expectancy, effort expectancy, Social Influence, and Facilitating Conditions. “Social Influence” refers to the degree to which an individual perceives that important others (such as friends, classmates, teachers) believe they should use the new technology. “Facilitating Conditions” refers to the individual’s perception of the organizational and technical resources available to support the use of the system. In recent years, numerous studies have applied UTAUT to explain the acceptance of AI chatbots or tools (Li et al., 2024; Qi et al., 2025).

For university students, these two new variables in the UTAUT model undoubtedly have strong explanatory power. For instance, recommendations from peers (Social Influence) or whether the university’s counseling center officially introduces and provides technical support (Facilitating Conditions) could significantly impact the adoption of AI music therapy.

Although the UTAUT model offers a more comprehensive explanatory framework, this study opts for a more parsimonious model. Considering that the core objective of this study is to deeply analyze the internal conflict between an individual’s psychopathological state (depression) and their perception of privacy risks, rather than external environmental factors, we have chosen a more streamlined model focused on core psychological variables. This study treats the core belief of the classic TAM (Perceived Usefulness) as an important control variable, thereby focusing the research on the non-rational emotional factor of depression. This allows for a clearer examination of the interplay between “deficiency compensation” and “privacy calculus.” We acknowledge this is a simplified perspective and suggest in the limitations section that future research should integrate the internal psychological mechanisms of this model with the external environmental factors of UTAUT to construct a more complete model of digital mental health intervention acceptance for university students (Yang et al., 2024).

With the development of big data technology, users face increasingly severe privacy risks when using digital services. The Privacy Calculus Theory, proposed by Culnan and Armstrong (1999), posits that individuals conduct a rational “cost–benefit analysis” when disclosing personal information (Xu et al., 2009).

Benefits: Include personalized services, convenience, economic rewards, or (in the context of this study) psychological comfort (Chaudhry and Debi, 2024).

Costs (Risks): Primarily the potential risks arising from privacy breaches, such as social discrimination (stigma), identity theft, or a sense of being monitored (Li et al., 2024).

The theory suggests that users will choose to relinquish their privacy only when the perceived benefits outweigh the perceived risks. In the context of AI music therapy, students face a typical privacy calculus trade-off: on one hand, they desire the emotional relief brought by AI (benefits), and on the other, they fear that the leakage of their depression data could lead to academic setbacks or social exclusion (costs). This trade-off may be even more intense in depressed populations, as they are more sensitive to negative evaluations.

The Deficiency Compensation Hypothesis provides the primary theoretical lens for H1. This hypothesis suggests that when individuals’ needs in real life (such as social interaction, emotional support) are not met, they tend to turn to online environments to seek alternative forms of satisfaction (McKenna and Bargh, 2014).

This hypothesis has been validated in multiple fields. For example, research shows that individuals with high levels of social anxiety are more inclined to self-disclose on the internet because the anonymous online space reduces their fear of negative evaluation (Sheldon, 2009). In the field of health communication, this effect is particularly pronounced. Patients with socially stigmatized diseases (such as HIV, schizophrenia) are more likely to seek information and emotional support in anonymous online health communities due to fear of real-world discrimination (Wright and Bell, 2003).

We extend this logic to the context of our study. For depressed university students, traditional face-to-face help-seeking is fraught with resistance, whereas AI music therapy, through its unique interactive features, provides precise “compensation” for these barriers (Potts et al., 2025). Specifically:

In the face of the fear of negative evaluation that traditional counseling might trigger, AI offers a completely non-judgmental interactive environment.

Addressing the concern of identity exposure due to stigma, the anonymous nature of AI provides a safe psychological space.

For individuals who find it difficult to initiate offline help-seeking due to social anxiety and energy depletion, the on-demand accessibility and depersonalized nature of AI tools significantly lower the barrier to seeking help.

In summary, these features of AI music therapy precisely meet the contradictory needs of depressed individuals who both crave help and fear interpersonal contact. Several empirical studies indirectly support this view, indicating that adolescents with higher levels of depression use the internet more frequently as an emotion regulation tool (Sun et al., 2025). Therefore, we have ample reason to infer that when real-world help-seeking channels are obstructed, the severity of depressive distress will act as a powerful driving force, prompting individuals to turn to this low-social-pressure, high-security digital form of help-seeking.

It is important, however, to acknowledge an alternative perspective. From a classic Technology Acceptance Model (TAM) viewpoint, depressive symptoms—often associated with cognitive fatigue, motivation, and reduced self-efficacy—could plausibly be interpreted as a factor that reduces perceived ease of use or increases perceived effort, thereby acting as a barrier to technology adoption. While acknowledging this possibility, our study focuses on the motivational aspect, conceptualizing depression as a ‘need’ that drives help-seeking in a low-stigma, non-interpersonal context. We hypothesize that for digital tools that bypass social barriers, the compensatory ‘pull’ of need will outweigh the ‘push’ of cognitive barriers, a premise we test with H1.

Based on this, we propose hypothesis H1:

Based on the logic of Privacy Calculus Theory (PCT), although depression may drive the intention to seek help, this process is not unconditional. Privacy Concern refers to an individual’s worries about the improper collection, use, or disclosure of personal information. In the digital health domain, health data (especially mental health data) is considered the most sensitive category of privacy (Kasen et al., 2024). For depressed students, the consequences of a data breach are not just nuisance calls but could lead to severe social exclusion and reputational damage.

Privacy Calculus Theory suggests that privacy concerns can inhibit the process of translating “need” into “action.” Existing research has shown that privacy concern is a significant factor influencing users’ adoption of AI chatbots for mental health (Li et al., 2024). Specifically:

Low Privacy Concern Context: When students perceive the AI system as secure or do not mind their data being collected, the distress from depression will directly translate into a strong motivation to seek help (consistent with the logic of H1). In this case, the more severe the depression, the higher the intention to use.

High Privacy Concern Context: When students are extremely worried about data leakage, this fear activates a “defense system.” Even if they are very depressed and in great need of help, the potential catastrophic consequences of a privacy breach (e.g., being advised to withdraw from school, being gossiped about by peers) will override the desire to seek help. At this point, the facilitating effect of depression on the intention to use will be offset by this “brake.”

That is, privacy concern is not just a direct negative factor but also a moderator that alters the strength of the relationship between depression and intention to use.

Based on this, we propose hypothesis H2:

The questionnaire consisted of three parts:

Informed consent and ethical statement: The first page of the questionnaire clearly informed participants of the research purpose, data anonymity, the voluntary nature of participation, and their freedom to withdraw at any time. It was specifically emphasized that all data would be used for academic statistical analysis only and would not involve personal identification, to minimize participants’ concerns.

Core Psychological Scales: Included measures for depression, privacy concerns, and technology acceptance.

Demographic Information: Included gender, age, grade level, major, and level of education.

The research subjects were university students (including undergraduates and postgraduates) in mainland China. Data was collected from March to April 2025. A combination of convenience sampling and snowball sampling was used to distribute the questionnaire link via online platforms like Wenjuanxing. To ensure sample diversity, the researchers distributed the questionnaire in student communities of comprehensive universities, science and engineering universities, and arts universities. To control data quality, the survey system had the following screening mechanisms, and manual cleaning was performed after data export:

IP Restriction: Each IP address was limited to one response to prevent duplicate submissions.

Patterned Response Detection: Invalid questionnaires with identical answers to all questions or showing obvious patterns were manually removed.

Response Time Control: Questionnaires completed in less than 90 s were excluded.

Considering that “AI music therapy” is a cutting-edge and diverse concept, to ensure all participants had a consistent understanding of the core research object, we provided a clear scenario description before the relevant sections of the questionnaire. This definition aimed to highlight its core features of “automation” and “personalization” while excluding the confounding factor of human intervention. The specific description was as follows:

[Scenario Description] Please read carefully:

Imagine that your university counseling center or an app store has launched an “AI Emotion Regulation Assistant.”

It does not involve any human counselors.

It analyzes your state (e.g., voice intonation, heart rate, or self-report) through an AI algorithm and automatically generates personalized soothing music for you.

The purpose is to help you relax when you are stressed or in a bad mood.

Through this standardized description, we operationally defined the research object as a personalized music generation service that requires no human interaction, is driven by algorithms, and aims at emotion regulation, thereby ensuring the validity of the construct measurement.

Depression Severity: Measured using the mature PHQ-8 (Patient Health Questionnaire-8) scale (Kroenke et al., 2009). This scale includes 8 items (e.g., “Feeling down, depressed, or hopeless”), scored from 0 to 3, and has good reliability and validity in non-clinical populations. It was used directly in this study without adaptation.

Privacy Concern: Adapted from the classic mobile internet privacy concern scale by Xu et al. (2009). The adaptation process was mainly contextualization: we specified the general term “personal information” in the original scale to “emotional data” and “mental health status data” in our research context. For example, the original item “I am concerned that my personal information might be used improperly” was adapted to “I am concerned that my mental health status data could be disclosed to the university or third parties.” This adaptation aimed to enhance the relevance of the items to the research context.

Intention to Use & Perceived Usefulness: Adapted from the corresponding scales in the UTAUT model by Venkatesh et al. (2003). The adaptation was also contextual, replacing “the system” in the original items with “this type of AI tool” or “this AI music tool.” For example, “I intend to use the system” was adapted to “If I have the opportunity, I am willing to try this AI music tool.” Similar adaptations are commonly used in other AI acceptance studies (Li et al., 2024; Qi et al., 2025).

To ensure the quality of the measurement tools, this study conducted reliability and validity tests.

Reliability Test: This study used Cronbach’s α coefficient to assess the internal consistency reliability of each scale. The specific results will be reported in Chapter 4.

Construct Validity Test: Since the classic Confirmatory Factor Analysis (CFA) has extremely strict model requirements, and considering the core purpose of this study was to validate the dimensional structure of the adapted scales, we employed Exploratory factor analysis (EFA) as a rigorous and widely accepted alternative to test the construct validity of the scales.

First, the data suitability test results showed a KMO value of 0.765 (greater than 0.7), and the Bartlett’s test of sphericity was significant (χ²(36) = 1991.84, p < 0.001), indicating that the data was very suitable for factor analysis.

Subsequently, we conducted an EFA on the 9 items of Perceived Usefulness, Privacy Concern, and Intention to Use, using Maximum Likelihood extraction and Varimax Rotation. The results clearly identified three factors, with the factor loading matrix shown in Table 2.

As shown in Table 2, the analysis results were perfectly consistent with the theoretical constructs: all items showed high loadings (>0.6) on their expected factors and extremely low cross-loadings on other factors. This indicates good discriminant validity among the three constructs. In summary, the EFA results provide strong evidence for the construct validity of the measurement tools used in this study.

Considering that gender and major background might influence attitudes toward new technologies, this study included them as control variables. Additionally, according to TAM theory, Perceived Usefulness is the strongest predictor of intention to use (Davis, 1989), so this variable was also measured and controlled.

As shown in Table 3, the Cronbach’s α coefficients for all scales exceeded the acceptable standard of 0.7. The internal consistency of the Depression scale (α = 0.888) and the Intention to Use scale (α = 0.875) reached excellent levels, indicating that the measurement tools used in this study have good reliability.

This study conducted descriptive statistical analysis on the core variables, with results shown in Table 4. It is noteworthy that the average total score for the sample on the PHQ-8 depression scale (0–24 point scale) was 6.07 (SD = 4.60). According to the standard clinical cutoffs for the PHQ-8, the average level of this university student sample falls within the “mild depression” range.

To test for data normality, this study combined numerical and graphical tests. Numerically, Table 4 shows that the absolute values of skewness and kurtosis for all variables are less than 1, which is well below the recommended threshold of 2. Graphically, the histograms in Figure 2 provide further visual evidence, showing that the distributions of all variables are well-shaped and approximate a normal distribution. In summary, the data distribution meets the prerequisite conditions for subsequent Pearson correlation analysis and multiple regression analysis.

The Pearson correlation coefficient matrix among the variables is shown in Table 5. To more intuitively display the relationships between variables, we further plotted a correlation heatmap (Figure 3).

The results clearly show that depression level is significantly positively correlated with intention to use (r = 0.270, p < 0.01), meaning that students with higher levels of depression have a stronger intention to use AI music therapy. Privacy concern is significantly negatively correlated with intention to use (r = −0.190, p < 0.01), indicating that students more concerned about privacy have a lower intention to use. Perceived usefulness has the strongest positive correlation with intention to use (r = 0.528, p < 0.01), which is highly consistent with the classic theory of the Technology Acceptance Model (TAM) (Davis, 1989) and also validates the effectiveness of the measurement tools in this study. These results provide strong preliminary evidence for the subsequent hypothesis testing.

Hypothesis H1 predicted that university students’ level of depression would positively influence their intention to use AI music therapy.

In Model 1 of Table 6, we first introduced the control variables (gender, major, perceived usefulness). The results showed that this model could explain 30.5% of the total variance in intention to use ( R 2 = 0.305).

Subsequently, in Model 2, we added the independent variable “Depression Level” to Model 1. The results showed that depression level had a significant positive predictive effect on intention to use (β = 0.128, p < 0.001). The introduction of this variable increased the model’s overall explanatory power from 30.5 to 32.0%, and the change in R² ( Δ R² = 0.015) was highly statistically significant (F-change = 11.21, p < 0.01). Furthermore, an examination of the model’s residuals indicated that they were evenly distributed around 0 with no obvious heteroscedasticity (see Supplementary material B), satisfying the basic assumptions of a linear regression model.

This positive relationship is intuitively demonstrated in the scatter plot in Figure 4: the overall trend of the data points indicates that as an individual’s depression score increases, their intention to use the AI music therapy tool also strengthens.

In summary, the regression analysis results provide strong support for Hypothesis H1.

Hypothesis H2 predicted that privacy concerns would play a negative moderating role in the relationship between depression and intention to use.

To test this hypothesis, we constructed Model 3, which included the interaction term. As shown in Table 6, after introducing the interaction term (Depression × Privacy Concern), its regression coefficient was significantly negative (β = −0.086, p < 0.05). This indicates that the level of privacy concern indeed significantly alters the strength of the effect of depression on intention to use.

To more clearly explain this interaction effect, we conducted a simple slope analysis and plotted the moderation effect (Figure 5).

Low Privacy Concern Group (−1 SD): As shown by the blue line, for students who were less concerned about privacy breaches, the positive predictive effect of depression level on intention to use was strong and significant (simple slope B = 0.041, p < 0.001). In this case, their emotional distress effectively translated into a motivation to seek AI-powered help.

High Privacy Concern Group (+1 SD): As shown by the red line, for students highly concerned about their privacy, the positive predictive effect of depression level on intention to use became flat and was no longer significant (simple slope B = 0.010, p = 0.312). This suggests that strong privacy concerns “inhibited” the help-seeking motivation; even if they had high levels of depressive mood, this mood did not effectively translate into an intention to use.

In summary, the results of the simple slope analysis clearly demonstrate the negative moderating effect of privacy concerns. Hypothesis H2 is fully supported.

The study found that after controlling for variables such as perceived usefulness, depression level still significantly and positively predicted university students’ intention to use AI music therapy (β = 0.128). This finding is consistent with the theoretical expectations of the ‘Deficiency Compensation Hypothesis,’ which posits that when individuals’ needs for emotional support are not met in reality, they may turn to low-social-pressure online environments for alternative satisfaction. Unlike previous research suggesting that depression leads to behavioral withdrawal, this study shows that when faced with non-interpersonal digital tools, depressed individuals are associated with a stronger approach motivation. This aligns with findings from some studies that students with mental health needs show a higher intention to use digital mental health solutions like apps and chatbots (Gbollie et al., 2023; Sun et al., 2025). The psychological mechanism behind this phenomenon may stem from the intrinsic nature of AI music therapy as a non-interpersonal interactive tool. Specifically, the ‘non-judgmental space’ it provides can precisely bypass the social evaluation and stigma concerns that may exist in traditional psychological help-seeking (Li et al., 2024). This is perhaps the key to its appeal to depressed individuals, and it also enables it to be an effective supplement to mental health services (Harith et al., 2022).

The most central finding of this study is the revelation of the significant moderating role of privacy concerns on the relationship between distress and help-seeking intention (H2). The interaction plot (Figure 5) clearly illustrates the “help-seeking vs. privacy” trade-off: in the low privacy concern group (blue line), the positive effect of depression on intention to use is very strong. However, in the high privacy concern group (red line), this positive pathway, while still present, is significantly weakened.

This statistical “chilling effect” provides a possible explanatory framework for understanding the real-world dilemma of why many depressed students, despite knowing they need help, hesitate to use mental health apps. For those depressed individuals with high privacy sensitivity, the fear of data leakage may override the need to seek comfort, thereby potentially being associated with a diminished intention to seek help. For depressed individuals, their psychological defenses are already fragile; they not only fear being spied on but also dread the social consequences of data leakage (such as being discriminated against). When “privacy anxiety” surpasses the “need for therapy,” they fall into a predicament of “the more pain, the more silence.” This is a specific manifestation of the “privacy paradox” in the mental health field (Barth and de Jong, 2017). It is also important to note that while the moderation effect is statistically significant and theoretically meaningful, its effect size is modest (β = −0.086). This is common in behavioral research where multiple factors influence intentions, but it underscores that privacy concern is one of several important factors in this complex decision-making process.