Concept
Psychology Education
Psychology education — the teaching of psychology as a discipline: the introductory and methods sequence, the undergraduate major, and the clinical and counseling training that leads to practice. It has the largest AI literature of any discipline in this base, and most of that literature does not study the psychology curriculum at all. Psychology students are the most convenient sample in higher education. They supply the respondents for generic studies of motivation and feedback. A smaller set of studies examines the courses themselves: assessment design in a second-year unit, an AI tutor in a statistics course, and the clinical training where AI is both a tool and a professional risk. The distinction matters because a finding that psychology students use generative AI in one way is not a finding about teaching psychology.
Questions to Consider
- Psychology students are the easiest sample to recruit in higher education, so most studies of "students and AI" are really studies of psychology students. What could a finding from a participant pool tell you about your own course, and what could it not tell you?
- A generated item bank can be reliable to acceptable standards while testing almost nothing above recall. Which matters more for the grade your students earn: that the marks are consistent, or that the items measure what the degree claims?
- The R-Tutor study prompts a model to guide rather than to answer inside a statistics course. If the tutor withholds the answer, what is it teaching that a chatbot that answers is not?
- Counselor trainees practiced with a simulated client that was agreeable, emotionally flat and culturally neutral. Which of those three would worry you most in a trainee's first real session, and why?
- Clinical training is where a tool becomes a professional claim, not a study aid: a system that recommends a treatment is making a judgment about a person. Where would you draw the line between the two, and who in your program should draw it?
Introduction
Psychology sits in an unusual position in this literature. It supplies the discipline's methods, its constructs and most of its respondents, and it also teaches a subject whose core skills — designing a study, reading a statistic, judging evidence — are the ones generative AI is most often accused of short-cutting. That combination makes psychology education both the best-documented discipline in the AI literature and one of the least studied as a curriculum.
Separating the two is the whole of the discipline's AI problem. What is known about teaching psychology itself is narrower than the volume of research suggests: assessment design in the undergraduate major, an AI tutor in the statistics sequence, the clinical and counseling training where AI arrives as a professional question, and the survey evidence on what psychology students know and believe. Studies that use psychology students as a convenience sample for a question about students in general are not evidence about the discipline, and the difference is worth naming each time.
Why psychology dominates the literature
Three conditions explain the volume. Psychology departments run participant pools, so recruiting a few hundred undergraduates to complete a survey costs almost nothing. The scholarship of teaching and learning is well established in the discipline, which gives instructors a route to publish course-level work that medicine and law mainly route through accreditation. And the constructs the field studies — motivation, self-efficacy, feedback, critical thinking — are the same constructs the AI literature wants to measure, which makes psychology the natural place to develop and validate instruments.
The result is a literature where the discipline is the setting rather than the subject. A survey of 991 undergraduates at five universities, 300 of them psychology majors, tells you how students in general use AI; it does not tell you whether a psychology degree should require a statistics course in which the tool is banned. Both questions are worth asking, and the second one is what the distinction is for.
The course-level evidence
The clearest course study is an assessment design. Richmond and Nicholls replaced peer assessment with ChatGPT in a second-year psychology "Media Assignment", where students translate a published article into a media release and video for a lay audience, and the redesign is about what the task asks for rather than whether the tool is allowed.
Assessment generation is the second strand. Baudin had ChatGPT-4 write 80 multiple-choice items for an Experimental Psychology course and ran the full psychometric apparatus over them with 120 students: internal consistency acceptable, test-retest moderate, 11.25% of items discriminating poorly, and a Bloom's distribution of 73.75% recall against 8.75% analysis or evaluation. The items were reliable evidence that students could name things, and silent about whether they could design a study.
Statistics is the third, and the best-evidenced intervention. Préau and colleagues deployed a purpose-built "R-Tutor" inside a semester-long R programming course for psychology undergraduates at Tübingen, in a preregistered longitudinal design, with the tutor prompted to guide rather than to answer.
Clinical and counseling training
Clinical psychology is where AI enters as a professional question rather than a study aid. The two studies here divide along that line: one on what a simulated client does to the trainee, one on how far trainees differ from each other.
Akkurt, Maurya and Brown used GPT-4o as a simulated client for counselor trainees. Nine transcripts and a four-person focus group is thin evidence, but the pattern is specific: the simulator lowered performance anxiety and built confidence, while the client agreed too readily, carried no emotional nuance and stayed culturally neutral — removing exactly the resistance and identity dynamics that counseling training exists to practice.
Armagan-Bogatekin modeled 286 psychology students and psychotherapy trainees across six domains and found three to four latent classes in every one. There is no average trainee: skepticism, learning anxiety, job anxiety and risk concern form separate profiles, and the profiles differ by whether the tool is a feedback aid or a treatment recommender.
What psychology students know, believe and do
The student-side evidence is the largest part of the literature and the least specific to the discipline.
Knowledge and judgment: Richter and colleagues had 150 psychology students classify 30 mental-health statements and compared them with four LLMs across 80 runs each; Reed and colleagues tested whether AI literacy predicts applied ethical judgment across 531 undergraduates with six vignettes.
Dispositions: Hefter, Paaßen and Berthold found critical GenAI use predicted by knowledge measured both by self-report and by objective test. Epistemic orientation and intellectual values carried the rest; Hu found continued use mediated by Trust and academic self-efficacy, with AI anxiety weakening the first link; Hawkins and colleagues found feedback literacy the only significant predictor of essay grade (β = 0.46) in a simulated AI-assisted writing task.
Use: Burke and colleagues surveyed 991 undergraduates across five universities and found college use at 81%, with studying displaced writing as the most common task, while 30.2% of students who call AI use cheating use it anyway.
Every one of those studies is about students who happen to be psychology students. That is the discipline's advantage as a research site and its weakness as a literature.
The psychology the research borrows
AI-in-education research runs on a small set of psychological theories that were built long before it. Naming them matters, because each predicts something different about whether a tool helps.
Cognitive offloading says a tool that removes effort removes the learning with it. Productive failure says the struggle before the answer is where the instruction happens. Scaffolding says help has to be scheduled and withdrawn, not simply provided. A study can cite all three and still design a tool that satisfies none of them, which is why the theory belongs in the methods section rather than the introduction.
The frames that recur most here are individual-cognition ones: self-regulated learning and Metacognition, self-efficacy, self-determination theory for motivation, and Feedback through its literacy variant. The studies in the section above are typical of the pattern: each takes a construct from this family, measures it as the mechanism, and reports which way the path runs. Akkurt is the exception, and the more interesting one: a situated design in which a trainee rehearses a practice rather than acquiring a proposition.
Two limits are worth stating. Theory is usually imported as a frame and rarely tested against a rival, the weakness theory development in AI education documents. And the individual unit of analysis is contested: sociocultural and activity-theory accounts, with distributed cognition, locate the learning in the room rather than the head, which changes what a chatbot is doing. Learning theories holds the map and Cognitive Psychology the machinery most of these studies assume.
When a paper says "scaffolding" or "critical thinking", ask which theory's version it means: the word does different work in each.
Implications for psychology instructors
- Decide what the degree is for before deciding what AI may do. A psychology graduate needs to evaluate evidence, design a study and interpret a statistic. If those are the outcomes, the assessment question is which tasks can still provide evidence of them.
- Use the statistics course as the test case. It has the strongest intervention evidence in the discipline (R-Tutor) and the clearest temptation to use AI as a calculator rather than a tutor.
- Audit generated item banks at the Bloom's level, not the reliability level. Baudin's bank was reliable and mostly recall.
- In clinical training, teach the tool and the boundary together. The trainee profiles differ, and the treatment-recommendation tool raises a professional-identity question that no capability training answers.
- Do not generalize from participant-pool samples to your own course. The generic findings describe students; your course is a design decision.
Common Misconceptions
- "Psychology students are representative of students in general." They are convenient, not representative, and the studies that use them usually say so.
- "AI in psychology education is mainly about clinical tools." The largest share of studies concerns writing, feedback and study habits; clinical tools are a small, recent and distinct strand.
- "Reliable generated items are usable items." Reliability without coverage of higher-order outcomes produces a well-measured test of the wrong thing.
- "Trainees are either for or against AI." The latent-class evidence says four or more profiles per domain, differing by tool.
Connected Concepts
Connected Articles
- Using generative AI to promote psychological, feedback, and artificial intelligence literacies in undergraduate psychology
- Assessing the psychometric properties of AI-generated multiple-choice exams in a psychology subject
- Student Adoption of an AI Tutor for Statistical Programming: A Longitudinal Study
- Learning Through Simulation: Counselor Trainees' Interactions with ChatGPT as a Client
- Heterogeneity in AI attitudes, anxiety, and acceptance among psychology students and psychotherapy trainees: a domain-specific latent class analysis
- Mental Health Literacy Across Psychology Students and Large Language Models
- AI literacy and scenario-based ethical judgment about generative AI among undergraduate students
- GenAI knowledge, epistemic orientation, and intellectual values predict undergraduate students' critical GenAI use
- Psychological predictors of continued ChatGPT use among university students: the roles of AI literacy, trust, and academic self-efficacy
- Summarise, elaborate, try again: exploring the effect of feedback literacy on AI-enhanced essay writing
- The Proliferation of AI among College Students: A Replication and 2-Year Follow-Up