Technical Summary: Perceptions of AI-CBT: Trust and Barriers in Chinese Postgrads

Problem Statement

The mental well-being of postgraduate students is a critical global concern, characterized by elevated rates of anxiety, depression, and suicidal ideation. In China, these risks are exacerbated by intense academic competition, heavy workloads, and high-stakes evaluations. Despite the severity of the issue, formal help-seeking remains low (approximately 2.7–3.1% in major cities) due to stigma and the perception that counseling is reserved for "serious problems."

While Artificial Intelligence-powered Cognitive Behavioral Therapy (AI-CBT) chatbots offer a scalable, low-barrier alternative with potential for 24/7 access and privacy, their adoption among Chinese graduate students remains uncertain. Existing literature on AI in mental health is limited by two factors:

1. Contextual Gap: Most data derives from clinical settings or general population surveys, failing to capture how specific high-stress student groups navigate AI-CBT within their unique cultural and institutional environments.
2. Temporal Gap: Many studies predate recent generative AI advances (e.g., more naturalistic dialogue), potentially misrepresenting current user expectations and concerns.

There is a lack of qualitative evidence regarding how Chinese graduate students evaluate the trust, effectiveness, emotional safety, and ethical implications of AI-CBT in their daily lives.

Methodology

This study employs a qualitative interpretive design to explore the perceptions and experiences of Chinese graduate students regarding AI-CBT chatbots.

• Participants: Ten Chinese postgraduate students (Master's or doctoral level) from Mainland China, Hong Kong, or Macau. Participants were recruited via WeChat and student networks using non-incentivized convenience sampling. Inclusion criteria required fluency in Chinese or English and familiarity with AI tools; exclusion criteria included current or past clinical mental health diagnoses to maintain a non-clinical focus.
• Data Collection: Semi-structured, one-on-one interviews were conducted via Zoom (or in-person), lasting 15–30 minutes. The interview guide was informed by the Health Belief Model (HBM) and the Theory of Planned Behavior (TPB), covering topics such as perceived usefulness, barriers, trust, emotional safety, data privacy, and social norms.
• Analysis: Data underwent reflexive thematic analysis following Braun and Clarke's six-phase framework. Coding was initially inductive, followed by a secondary interpretive pass using HBM and TPB as sensitizing frameworks to map findings onto constructs like perceived susceptibility, barriers, attitudes, subjective norms, and perceived behavioral control.
• Rigor: Methodological integrity was ensured through an audit trail, reflexive journaling, peer debriefing, and the pursuit of thematic sufficiency rather than statistical generalizability.

Key Findings

The analysis yielded five interrelated themes that characterize a "cautious openness" toward AI-CBT:

1. Perceived Usefulness and Accessibility ("It's There When I Need It"): Participants valued the 24/7 availability, immediacy, and anonymity of AI-CBT. It was viewed as a low-stigma resource for managing everyday stressors and "venting" late at night, rather than a substitute for intensive therapy.
2. Limitations of Empathy and Trust ("It's Not a Real Person"): A significant barrier was the perception that chatbots lack genuine emotional depth. Responses were often described as overly rational, repetitive, or generic, leading to skepticism about the tool's ability to handle complex emotional issues.
3. Privacy and Ethical Concerns ("Is My Data Safe?"): Data security was a primary deterrent. Participants frequently self-censored sensitive disclosures due to uncertainty about how data would be used by institutions or companies. This directly limited the perceived usefulness and willingness to engage deeply.
4. Social Norms and Cultural Stigma ("It's for Others, Not for Me"): Cultural stigma surrounding mental health influenced adoption. Many participants distanced themselves from using AI-CBT, viewing it as a tool for "others" (e.g., the lonely or introverted) to avoid peer judgment or reputational damage.
5. Conditions for Adoption and Self-Efficacy ("I'd Try it, But Only If..."): Adoption was conditional. Students expressed willingness to use AI-CBT only if specific safeguards were met: transparent data handling, more natural/empathetic interactions, and culturally relevant design. They viewed the tool as a complementary aid for surface-level issues, not a replacement for professional care.

Theoretical Integration

The study integrates HBM and TPB to explain these findings:

• HBM: Students weigh perceived benefits (accessibility, privacy, immediacy) against barriers (lack of empathy, data risks, stigma). Signals to action (e.g., stress, loneliness) are often overridden by these barriers.
• TPB: Positive attitudes toward the utility of AI-CBT are constrained by negative subjective norms (peer stigma) and reduced perceived behavioral control (distrust in the technology's capability and safety).

Contributions

The paper makes three specific contributions:

1. Empirical: It provides a first-person account of AI-CBT perceptions among an underrepresented, high-stress cohort (Chinese graduate students), moving beyond general population surveys.
2. Conceptual: It demonstrates how HBM and TPB can be integrated to interpret qualitative data, clarifying how benefits, barriers, norms, and control co-produce adoption intentions in a non-clinical academic setting.
3. Practical: It outlines design and deployment implications, emphasizing the need for transparent data handling, emotional safety safeguards, culturally tuned onboarding, and clear escalation pathways.

Significance and Claims

The authors claim that the study offers context-specific insights to guide the culturally sensitive design and communication of AI mental well-being tools for student populations in China. The significance lies in identifying that while there is a demand for low-barrier support, trust is the critical bottleneck.

The paper asserts that sustained engagement with AI-CBT depends on three factors:

1. Building trust through transparent data practices and explainable responses.
2. Strengthening perceived control through clear guidance and safety assurances.
3. Normalizing use within academic communities to reduce stigma.

The authors explicitly state that their findings are not intended for statistical generalization due to the small, non-probability sample. Instead, the study serves as a foundation for future work to test these hypotheses with larger, diverse samples and to experimentally evaluate specific design characteristics (e.g., tone, privacy disclosures) on trust and intention.