GenAI as a Frenemy in Teaching: Perceived Autonomy and Risks

Synthesis: Dai, Ni, Meng, Ju, Crawford, and Teo (2026) use a human-AI interaction lens to ask why teachers vary in their willingness to adopt generative AI, combining a 287-participant, 27-country survey with structural equation modeling and thematic analysis of 285 open-text responses. Extending the technology acceptance model with perceived artificial autonomy (AA) and risk aversion (RA), they find that teachers see GenAI as only semi-autonomous - most chose "teacher assistance" or "partial automation" on a six-level model - yielding a refined model that explains 79.8% of the variance in behavioral intention. AA influenced intention only indirectly through perceived usefulness (indirect effect .156, 95% CI [.082, .240]); its direct path was non-significant (.026), because adoption hinged on situational, pedagogical judgment rather than the appeal of automation. Risk aversion was a significant negative predictor of intention (-.163), and teachers worried far more about students' GenAI use - shortcuts, overreliance, hallucinations, and integrity risks - than about their own. The authors frame GenAI as a "frenemy": a valuable pedagogical resource whose limited autonomy and perceived risks keep humans firmly in control, implying that adoption depends less on technical capability than on context-sensitive support, clear policy, and professional development.

Key Findings

Teachers positioned GenAI as semi-autonomous, not autonomous: On the six-level automation model, most chose mid-range levels - 130 selected "teacher assistance" (Level 2) and 94 "partial automation" (Level 3) - against only 12 who answered "teacher only" (Level 1) and a single respondent endorsing full automation (Level 6).
The refined model explained 79.8% of the variance in behavioral intention: Perceived usefulness was the dominant predictor of intention (.832), and perceived ease of use drove usefulness (.479), confirming the core TAM pathways after attitude was removed.
Artificial autonomy operated only indirectly: The direct effect of AA on behavioral intention was non-significant (.026, 95% CI [-.028, .081]), while the indirect effect through perceived usefulness was significant (.156, 95% CI [.082, .240]) and the total effect positive (.182, 95% CI [.094, .273]).
Risk aversion suppressed adoption intention: Risk aversion had a significant negative association with behavioral intention (-.163, p < .001) but no significant link to perceived usefulness (-.063), showing that cautious teachers may still value GenAI while hesitating to adopt it.
Attitude was dropped for poor discriminant validity: Correlations between attitude and behavioral intention (.929) and attitude and perceived usefulness (.931) exceeded the square roots of their AVEs; the refined structural model fit well (chi-square(97) = 202.597, chi-square/df = 2.089, CFI = .963, TLI = .954, RMSEA = .062, SRMR = .041).
Prior experience changed the ease-of-use relationship: PEU to intention was non-significant in the full sample (.068), but a post hoc analysis of the 225 teachers who had previously used AI for teaching found it significant (.138, p < .05), with 78.7% of intention variance explained.
Risks were perceived as greater for learners than for teachers: Thematic analysis of 285 responses (17,988 words) produced 47 codes, narrowed to 22 relevant codes and two themes; 15 teachers reported no concerns, and the risks named clustered around student cheating, weakened foundational and higher-order thinking, hallucinations, reduced human interaction, copyright, data privacy, and bias.
Adoption was conditional on context, not autonomy: Teachers described case-by-case decisions weighing student needs, disciplinary differences, task type, intended user and timing, echoing one participant's view that GenAI "can only be used to assist and not complete it in its entirety".

Study Design & Method

Explanatory mixed methods design, with a quantitative questionnaire analyzed first and qualitative open-ended responses used to interpret and extend the results.
287 teachers from 27 countries and regions after screening; from 319 respondents, 32 were excluded (6 non-teachers, 16 without GenAI experience, 10 unengaged responses flagged by a reverse-coded item and zero variance).
Sample composition: 132 male and 155 female; 147 based in Europe, 76 in North America, 51 in Asia; 166 assistant professors or lecturers, 56 associate professors and 36 professors; 142 held doctorates; disciplines spanned arts and humanities (69), social sciences (70), natural sciences and mathematics (37), education (34), business and law (27) and ICT (19).
Measures: 20 seven-point Likert items covering five constructs (behavioral intention, attitude, perceived usefulness, perceived ease of use, risk aversion) adapted from established scales, plus a visual six-level automation scale for perceived artificial autonomy and two sets of open-ended questions; ChatGPT served as a familiar proxy for GenAI, and two educational-technology experts reviewed face and content validity.
Data collection and analysis: questionnaire distributed via Prolific between September 2023 and May 2024 (roughly 2.40 GBP compensation per participant); a two-step structural equation modeling approach in Mplus 7 (confirmatory factor analysis then path analysis with a robust maximum likelihood estimator, since multivariate normality was not supported), bootstrapped mediation tests with 95% confidence intervals, and Braun and Clarke thematic analysis in NVivo 14.
Limitations: self-reported, single-source data; an online sample that likely skews toward digitally literate and more favorable teachers; and a cross-sectional design that cannot track adoption over time.

What this means for practice

Teachers. Place each GenAI task on a shared automation scale before adopting it and state the level you settle on — 130 of 287 surveyed teachers chose "teacher assistance" and 94 "partial automation", against a single respondent endorsing full automation.
Institutions. Sequence professional development by experience: perceived ease of use predicted intention only among the 225 teachers who had already used AI in their teaching (β = .138), while the full-sample path was non-significant (.068).
Policymakers. Publish explicit acceptable-use expectations rather than reassurance, because risk aversion reduced adoption intention (-.163, p < .001) without reducing perceived usefulness (-.063).
Educators. Keep two judgments separate in local decisions: "useful" is not "safe to delegate", since perceived usefulness dominated intention (.832) while perceived autonomy had no direct effect on it (.026).
Teachers. Meet student-facing risks with task design: the 285 open-text responses clustered on cheating, weakened foundational and higher-order thinking, hallucinations, reduced human interaction, copyright, privacy and bias.

Limitations

Web-based panel of prior users. Only teachers with prior GenAI experience were allowed to participate (319 respondents screened down to 287, recruited through Prolific at about £2.40 each), so the model describes teachers already inclined to try it and says nothing about first-time users.
Single-source, cross-sectional self-report. One questionnaire supplies both the predictors and behavioral intention, and data were collected between September 2023 and May 2024, so adoption cannot be tracked over time.
Online sample skew. The authors state that respondents completing the survey online may have higher digital literacy and more favorable perceptions of GenAI than other teacher groups.
A construct was dropped for discriminant validity. Attitude had to be removed from the model because its correlations with behavioral intention (.929) and perceived usefulness (.931) exceeded the square roots of their AVEs, so the 79.8% explained variance is a property of the trimmed model.

Connected Concepts

Connected Articles

Factors Associated with Students' Adoption of Artificial Intelligence Technology in Tertiary Education: A Meta-Analytic Review - TAM-based meta-analysis of AI adoption in higher education
Factors influencing university students' intention to use and reliance on generative artificial intelligence - Technology acceptance model applied to critical GenAI use
Mind the Trust Gap: Identifying (Mis)alignments in Teacher-Student Views Toward Control and Agency in K-12 Classroom AI - Teacher and student views on control and agency in AI classrooms
Guarded adoption of generative AI in higher education: high-achieving students, successful-student identity - Cautious, risk-aware adoption of GenAI in higher education
AI Adoption Among Teachers: Insights on Concerns, Support, Confidence, and Attitudes - Teacher confidence as a driver of AI adoption

Citation

Dai, H. M., Ni, K., Meng, H., Ju, B., Crawford, J., & Teo, T. (2026). GenAI as a frenemy in teaching: Perceived autonomy and risks. Australasian Journal of Educational Technology, 42(3), 157-179.

Metadata

Study design	Survey, Thematic Analysis
Level	Higher Ed
Intended audience	Instructors, Institutions, Policymakers
Technologies and techniques	Generative AI, Technology Adoption Models
Research methods and evaluation	Mixed-Methods Research
Foundations of AI in education	Academic Integrity, Learner Agency, Cognitive Offloading, Educational Development, Human AI Collaboration, Teaching
Page kind	Framework

Source History Edit Chat