Research Article
Heterogeneity in AI attitudes, anxiety, and acceptance among psychology students and psychotherapy trainees: a domain-specific latent class analysis
Synthesis: Armagan-Bogatekin (2026) modeled 286 psychology students' and psychotherapy trainees' orientations toward two AI clinical tools — an automated feedback tool and a treatment-recommendation tool — with six separate latent class models. Every domain resolved to three or four classes, and the largest class was usually the moderate one. There is no average trainee: skepticism, learning anxiety, job anxiety and risk concern each form their own profile, and the profiles differ by domain, which means "AI acceptance" is not one attribute to train.
Key Findings
- No single profile describes the sample. Across all six domains, three- to four-class solutions fit better than a one-class description of a uniform trainee population.
- Moderation dominates. In several domains the largest class was the moderate or mixed pattern rather than a strongly favorable or strongly unfavourable one, which complicates the usual "students are enthusiastic" headline.
- Anxiety splits into job-related and learning-related. These emerged as distinct classes rather than as one anxiety dimension, and they sat differently against the two tools.
- Domain specificity is the main result. Because each domain was modeled separately, the classes describe how a trainee orients toward a feedback tool or a recommendation tool, not a general disposition toward AI.
- Risk and autonomy concerns form their own group. A class characterized by heightened concern about AI risk appears alongside the skeptical and anxious classes, which is a distinct position rather than the tail of a scale.
What was modeled
The sample is 286 psychology students and psychotherapy trainees, mostly in Türkiye, responding to measures of general AI attitudes, technology acceptance, readiness, job-related anxiety, learning anxiety, and concerns about AI risk and autonomy. Latent class analysis was run separately for each domain, which is the design decision that produces the paper's central claim: the domains do not collapse into one profile.
The theory framing draws on the Unified Theory of Acceptance and Use of Technology, technostress, and professional-identity-threat accounts. Those three literatures predict different things — acceptance, strain and threat — and the class structure suggests each is capturing a real subgroup rather than a measurement artifact.
Why a training program cannot be one-size
The practical implication follows directly from the class structure. A single induction session on AI in clinical training addresses the moderate class, alienates the skeptical one, and misses the anxious one, because each needs something different: the skeptical trainee needs evidence, the learning-anxious trainee needs scaffolded practice with a way to stop, the job-anxious trainee needs a clear account of what the tool does not replace, and the risk-concerned trainee needs governance answers rather than capability training.
The professional-identity frame is the one to watch in clinical education specifically. A treatment-recommendation tool is not a study aid; it is a claim about clinical judgment, and trainees who read it that way are not being irrational. The paper's implicit argument is that programs should teach the tool and the boundary at the same time, and expect the boundary conversation to differ by subgroup.
What this means for practice
- Clinical educators. segment by orientation before designing training. The four profiles need different content, not different amounts of the same content.
- Instructors. separate learning anxiety from job anxiety in the messaging. The first responds to scaffolding, the second to a credible account of professional role.
- Curriculum designers. teach the automated feedback tool and the recommendation tool as different objects. The domains did not merge in the data.
- Programmes. treat skepticism as a position with content, and answer it with validation evidence rather than reassurance.
Limitations
Cross-sectional, self-report, secondary-data design, as the author states; the classes are response patterns, not verified dispositions. The sample is 286 trainees in one national context, and class enumeration with six models raises the usual questions about replicability across samples. Domain-specific modeling forfeits the ability to describe a single integrated profile, which the author acknowledges as a deliberate trade. Item wording for AI-supported clinical tools may not transfer to tools trainees have not used.
Citation
Armagan-Bogatekin, M. (2026). Heterogeneity in AI attitudes, anxiety, and acceptance among psychology students and psychotherapy trainees: a domain-specific latent class analysis. Frontiers in Psychology, 17, 1859806.