Optimists, Neutrals, Skeptics: Study Maps How Students Really Feel About AI

Generative artificial intelligence has swept through university campuses faster than almost any educational technology before it, yet a new study suggests that beneath the apparent uniformity of ChatGPT-wielding students lies a strikingly divided landscape of opinion. Researchers at the University of Texas at Arlington and Penn State University surveyed 237 U.S. university students across STEM and non-STEM fields and, using a two-stage statistical clustering approach, found that students fall into three distinct perception profiles: optimists, who make up 41.4 percent of respondents; neutrals, at 37.6 percent; and skeptics, accounting for 21.1 percent. The findings, published in SN Social Sciences, challenge the assumption that students hold a single, monolithic attitude toward AI in education.

The study’s methodology was designed to move beyond simple averages. Rather than asking whether students like or dislike AI in the aggregate, the team administered a 25-item questionnaire covering perceived challenges, perceived benefits, and perceived effectiveness of generative AI tools. After exploratory factor analysis and reliability testing—during which two challenge items failed to align with the empirical structure and were removed—23 items remained. The researchers then applied hierarchical clustering using Ward’s method, followed by K-means refinement, to sort respondents into homogeneous groups based on their full pattern of responses.

The statistical case for three profiles was compelling. The agglomeration coefficient increased by only about 6.4 percent when moving from four clusters to three, but jumped by roughly 21.1 percent when collapsing from three clusters to two, indicating that merging below three groups would sacrifice substantial within-group homogeneity. Although a two-cluster solution produced a higher silhouette coefficient, it lumped 199 of 237 respondents into a single cluster, obscuring meaningful differences. Crucially, the researchers validated their clusters using a separate four-item ethical-concern scale that had been deliberately excluded from the clustering process, and the three profiles differed significantly on this external measure—skeptics scoring highest at 4.775 on a five-point scale, optimists lowest at 3.875.

The profiles themselves tell vividly different stories. Optimists reported the highest perceived benefits and effectiveness, strongly agreeing that AI provides personalized feedback that improves learning (mean 4.30), enhances access to resources (4.19), and speeds up understanding of concepts (4.17), while reporting comparatively low concern about reduced creative thinking (2.34) or ethical risks (2.53). Neutrals occupied a middle ground, acknowledging practical advantages like improved resource access while maintaining moderate concerns about ethics and emotional disengagement. Skeptics, by contrast, showed minimal endorsement of nearly every benefit item—their mean agreement that AI increases academic engagement sat at just 1.24—and the strongest concerns, rating ethical worries at 4.82 and agreeing that AI limits creative thinking at 4.30.

Perhaps the most eye-catching finding concerns what separates these groups. Frequency of AI use emerged as the single strongest correlate of profile membership, and the association was dramatic: among students who said they always used AI, 87.9 percent were optimists and not a single one was a skeptic. Among those who never used AI, 93.8 percent were skeptics. In multinomial logistic regression that simultaneously accounted for gender, field of study, educational status, and tech-savviness, each step toward less frequent AI use reduced the odds of optimist membership by roughly 77 percent and increased the odds of skeptic membership nearly sevenfold. The relationship held across 22 of the 23 individual perception items after correction for multiple testing.

Gender provided a second, independent signal. Male respondents had about 2.45 times greater odds than female respondents of belonging to the optimist rather than the neutral profile, and gender was associated with 21 of the 23 perception items, with the strongest single association involving ethical concerns about AI use. Field of study showed significant bivariate associations—STEM students were more likely to be optimists (57.3 percent) than their non-STEM peers (28.3 percent)—but this difference disappeared in the adjusted model, suggesting that STEM students’ sunnier outlook may reflect greater AI exposure rather than disciplinary culture per se. Tech-savviness and educational status, notably, did not independently predict profile membership, and none of their item-level associations survived false-discovery-rate correction.

Tool usage itself was highly concentrated. ChatGPT dominated, reported by 84.8 percent of students, followed by Grammarly at 48.1 percent and Quizlet at 38.0 percent. Turnitin and Google Scholar AI appeared in the second tier, suggesting that many students encounter AI not only as a content generator but through academic integrity systems and research-discovery tools. Usage fell sharply for Duolingo, Khan Academy’s AI, and Socratic, each reported by fewer than 15 percent of respondents.

The authors are careful to flag a critical caveat: perceived effectiveness is not the same as actual learning. Optimists’ strong belief that AI improves their grades and learning does not demonstrate that it does. Indeed, cited evidence includes a study of high school mathematics in which unrestricted GPT-4 access improved performance during AI-assisted practice but reduced subsequent performance on unassisted exams—a reminder that efficiency gains with AI support may not translate into durable independent skill. The cross-sectional design also means causality runs both ways: frequent use may breed optimism, or optimistic students may simply use AI more often.

The practical implications are aimed squarely at universities struggling to write coherent AI policies. The researchers argue that a one-size-fits-all approach is likely to fail and propose differentiated strategies: for optimists, training in verification, critical evaluation of outputs, and clear academic integrity expectations to guard against overreliance; for neutrals, concrete examples of appropriate AI-supported workflows, such as feedback-based revision and concept checking, paired with honest discussion of hallucinations and inconsistent outputs; and for skeptics, trust-building through transparent policies, explicit privacy protections, and optional, low-risk activities that let hesitant students evaluate AI on their own terms without making it a requirement.

What emerges is a portrait of higher education in which generative AI is neither savior nor menace but a mirror reflecting students’ varied experiences, trust levels, and disciplinary habits. With roughly four in ten students embracing AI enthusiastically, a similar share hovering in cautious ambivalence, and one in five remaining deeply skeptical, the study suggests that the campus debate over AI is far from settled—and that institutions hoping to navigate it will need to speak to three very different audiences at once.

Subject of Research: University students' perception profiles of generative AI in STEM and non-STEM higher education

Article Title: Cluster-based analysis of university students’ perceptions of generative AI across STEM and non-STEM fields

Article References: Zohourian, M., Pamidimukkala, A., & Kermanshachi, S. (2026). Cluster-based analysis of university students’ perceptions of generative AI across STEM and non-STEM fields. SN Social Sciences, 6(10), Article 467. https://doi.org/10.1007/s43545-026-01761-6

Image Credits: AI Generated

DOI: 10.1007/s43545-026-01761-6

Keywords: generative AI, higher education, student perceptions, cluster analysis, STEM education, ChatGPT, AI ethics, educational technology, multinomial logistic regression, AI-use frequency, academic integrity, person-centered analysis