Hidden Architecture of AI Acceptance: New Study Reveals How Students Really Embrace Generative AI

Generative artificial intelligence has swept through university campuses faster than almost any technology in the history of higher education, and researchers are still racing to understand why some students embrace it wholeheartedly while others hesitate. A new study published in Current Psychology by Ömer Faruk Ursavaş of Karadeniz Technical University and colleagues, including Hatice Yildiz Durak, İlknur Reisoğlu, Yavuz Akbulut, and David Mcilroy, tackles a deceptively simple question with unusually sophisticated statistical machinery: how should we actually measure whether students accept generative AI? The answer, the team argues, is that acceptance is not a flat collection of independent attitudes but a layered, hierarchical construct, and the tools psychologists use to measure it must reflect that complexity.

The study focuses on the GenAI-TAM scale, a 46-item instrument built on the Technology Acceptance Model, the influential framework first formalized by Fred Davis in 1989. The original model proposed that whether people adopt a technology depends largely on two perceptions: how useful they believe it to be and how easy they find it to use. Over the decades the model has grown, spawning revised versions and unified theories that add social influence, facilitating conditions, and other ingredients. The GenAI-TAM scale stretches this tradition across ten factors, capturing facets of acceptance that range from perceived usefulness and ease of use to attitudes, behavioral intentions, and the anxieties and social norms that swirl around tools like ChatGPT. But a scale with ten separate factors raises an immediate measurement problem: are those ten dimensions genuinely distinct, or do they all draw from one underlying disposition to accept generative AI?

Traditional validation methods struggle with exactly this question. Confirmatory factor analysis, the workhorse technique of psychometrics, forces every survey item to load onto one and only one predetermined factor, with all other relationships fixed at zero. That constraint works reasonably well for cleanly separated constructs, but human attitudes are messy. A student who believes ChatGPT is useful probably also finds it enjoyable, feels social pressure to use it, and intends to keep using it, and those attitudes bleed into one another. When items are allowed only a single loading, the model absorbs that cross-talk as error, and fit statistics deteriorate. The research team suspected that this was precisely what was happening with existing measures of generative AI acceptance, and they designed their study to find out.

Their approach was a systematic model comparison. Rather than testing a single favored model, the researchers specified seven competing structures and let the data arbitrate among them. At one end of the spectrum sat the traditional confirmatory factor analysis with its rigid simple structure. At the other end sat increasingly flexible frameworks drawn from the exploratory structural equation modeling tradition, a family of techniques introduced by Tihomir Asparouhov and Bengt Muthén in 2009 that blends the confirmatory power of theory-driven modeling with the freedom of exploratory factor rotation. In ESEM, items are permitted small secondary loadings on factors other than their primary one, acknowledging that attitudes naturally correlate across dimensions. The most elaborate candidate was a bifactor ESEM model, which simultaneously estimates one general factor running through all 46 items and ten specific factors capturing what remains unique to each dimension.

The data came from 860 undergraduate students, and the verdict was unambiguous. The confirmatory factor analysis models proved inadequate, failing to reproduce the patterns in the students’ responses. The bifactor ESEM model, by contrast, fit the data excellently, with a comparative fit index of .972, comfortably above the .95 benchmark that methodologists such as Liang Hu and Peter Bentler consider evidence of close fit, and a root mean square error of approximation of .046, well under the .06 threshold for acceptable error. In plain terms, the hierarchical model captured the structure of student attitudes with remarkable fidelity, while the simpler structures left systematic misfit behind.

The bifactor results carried a second, arguably more consequential finding. The general factor, which the researchers label GenAI Acceptance, showed a McDonald’s omega hierarchical coefficient of 0.87, a measure of how reliably the total set of items reflects the single overarching construct. Values above roughly 0.80 indicate that the common factor dominates the response patterns, and 0.87 is a strong result. This means the scale can legitimately be summarized with a single total score of generative AI acceptance, even though its items were designed to tap ten distinct facets. The specific factors still matter, because they describe where in the acceptance landscape a given student’s responses diverge from the general trend, but the dominant signal is the general one.

The team also tested measurement invariance across gender, asking whether the scale measures the same construct in the same way for male and female students. This step is often skipped, yet it is essential for fair comparisons. If an instrument functions differently across groups, a higher average score in one group might reflect measurement bias rather than a genuine difference in acceptance. The analyses confirmed that the bifactor ESEM model held across gender, and scalar invariance was supported, meaning that latent mean scores can be compared meaningfully between male and female students. Any observed gender differences in acceptance, in other words, would reflect real differences in attitudes rather than artifacts of the instrument.

The practical implications reach well beyond psychometrics. Universities are currently drafting policies on generative AI, deciding when it may be used in coursework, how academic integrity should be redefined, and what training students need. Sound decisions require sound measurement, and administrators who want to survey their student populations now have a validated instrument with a defensible structure. Researchers studying the antecedents of acceptance, from self-regulated learning skills to AI anxiety and trust, can use the GenAI-TAM scale knowing that its factor structure has survived a rigorous stress test. The finding that acceptance behaves as a hierarchical construct also carries theoretical weight, suggesting that models of technology acceptance should accommodate a general acceptance disposition alongside specific perceptual dimensions, rather than treating each facet as an isolated variable.

The authors are careful about the limits of their evidence. The validation establishes the internal structure, reliability, and gender invariance of the scale, but they note that further criterion-related and predictive validity evidence is needed before the instrument can be used for diagnostic profiling or institutional decision-making. In other words, showing that the scale measures a coherent construct is a necessary first step, but demonstrating that its scores predict actual usage behavior, academic outcomes, or responses to interventions requires future longitudinal and applied studies. The sample of 860 undergraduates, while substantial, also invites replication across cultures, disciplines, and educational levels, particularly given that prior comparative work has documented differences in generative AI acceptance between countries such as Poland and Egypt and across fields of study.

What makes the study resonate beyond its immediate field is the broader lesson about measurement in the age of AI. As generative tools become woven into writing, coding, research, and creative work, the constructs psychologists use to describe our relationship with them grow more intricate. Simple structures, the study shows, can be actively misleading: forcing multidimensional attitudes into rigid models produces poor fit and potentially distorted conclusions. The bifactor ESEM framework offers a way to honor both the unity and the diversity of human attitudes, revealing that student acceptance of generative AI is at once a single powerful disposition and a rich tapestry of specific perceptions. For educators and researchers navigating the fastest technological shift higher education has seen, that nuance may prove to be the difference between policies built on guesswork and policies built on evidence.

Subject of Research: Psychometric validation of the GenAI-TAM scale measuring undergraduate acceptance of generative AI using bifactor exploratory structural equation modeling

Article Title: Beyond simple structures: a bifactor-ESEM Approach to understanding student acceptance of generative AI in higher education

Article References: Ursavaş, Ö. F., Yildiz Durak, H., Rei̇soğlu, İ., Akbulut, Y., & Mcilroy, D. (2026). Beyond simple structures: a bifactor-ESEM Approach to understanding student acceptance of generative AI in higher education. Current Psychology, 45(17), Article 1432. https://doi.org/10.1007/s12144-026-09959-w

Image Credits: AI Generated

DOI: 10.1007/s12144-026-09959-w

Keywords: generative AI, technology acceptance model, higher education, psychometrics, bifactor ESEM, measurement invariance, scale validation, undergraduate students, ChatGPT, factor analysis, Current Psychology, educational technology