The research team, led by Zhenzhen Huang of the university library and colleagues at the School of Computer Science and Technology and the Chinese Academy of Agricultural Sciences, set out to resolve a genuine standoff in the literature. Survey-based studies have reported both that students increasingly treat AI as their primary information source and that AI tools help them find and understand literature more efficiently, encouraging further use of academic resources. The problem, the authors argue, is that nearly all of this evidence comes from self-reported questionnaires, which are notoriously poor proxies for what people actually do.

The study’s central innovation is a distinction between two very different ways students use generative AI. Auxiliary use means deploying AI to search for, screen, summarize, and explain academic material while still reading and synthesizing the sources oneself. Substitutive use means asking AI to generate complete passages for papers, reports, or assignments, bypassing the reading entirely. These two behaviors carry opposite theoretical implications: the first should complement library resources, the second should displace them. Prior research had largely lumped them together.

To test the two paths separately, the researchers built a structural equation model from the survey responses, measuring constructs such as deep-reading demand, perceived convenience of digital resources, and continuance intention toward print and digital resources. The results were striking. Deep-reading demand showed a strong positive association with intention to keep using print books, with a standardized coefficient of 0.908 and a large effect size. Auxiliary AI use was positively associated with intention to keep using digital resources, at a coefficient of 0.844. But auxiliary use showed no significant link to print-resource intention, and substitutive use showed only a weak negative association with print resources, a coefficient of just minus 0.100, and no significant link to digital resources at all.

The measurement work also surfaced an important caveat. Perceived convenience of digital resources and auxiliary AI use were so highly correlated in the sample, with a latent correlation of 0.967, that the two constructs failed standard discriminant-validity tests. The authors are candid about this: the near-perfect association may partly reflect common method bias, and the impressive-looking coefficients on the digital channel should be read as associations between overlapping constructs rather than as evidence of a clean psychological mechanism. Four other constructs were measured with single items, further limiting interpretation.

The behavioral data tell an equally nuanced story. Print borrowing at the university has indeed collapsed, falling from 406,120 loans in 2017 to 46,339 in 2025, with unique borrowers dropping from over 30,000 to under 7,300. But an interrupted time-series analysis with a prespecified breakpoint at January 2023, following ChatGPT’s release, found no significant level change and no significant slope change after that date. The long decline began years before generative AI arrived, and it did not accelerate when AI did. Alternative breakpoints in March 2023, September 2023, and January 2025 produced the same null result.

Digital-platform statistics reinforce the coexistence picture. Traditional operations on the library’s digital resource platform, including search, browse, and download, actually rose from about 280 million in 2021 to 317 million in 2025. Platform-embedded AI features, such as AI search and AI-assisted reading, first appeared in 2025 and grew quickly, reaching 766,000 operations in the first half of 2026. Yet that still represents only 0.53 percent of total platform activity. AI features are being adopted as a new layer within existing digital services, not diverting users away from them.

Triangulating the two evidence streams, the researchers found that the survey-level substitution signal, the weak negative association between substitutive AI use and print-resource intention, has not yet shown up in aggregate borrowing behavior. They offer two possible explanations. One is an attitude-behavior gap: students may overestimate how much AI has changed what they actually do. The other is a time lag: attitudinal shifts may precede behavioral ones, and micro-level substitution may not yet have accumulated enough to register in institutional data. Either way, the authors stress that the pre-existing decline in print borrowing, driven by forces predating ChatGPT, remains the more pressing concern for libraries.

A hypothesized moderating effect of task-channel fit, the idea that students who match AI tools to appropriate tasks would benefit more, was not supported, suggesting that how students divide their use across print, digital, and AI channels may reflect relatively stable habits rather than deliberate task matching. The authors frame their findings through self-regulated learning theory and cognitive offloading: students delegate different stages of information work to AI depending on the task, and auxiliary delegation appears compatible with continued engagement with academic resources.

The study’s limitations are clearly acknowledged. It is a cross-sectional survey at a single research-intensive Chinese university, the behavioral data are aggregate and cannot be linked to individual respondents, and the interrupted time-series design can only test whether the post-2023 period coincides with a trend change, not establish causation. The pandemic and its aftermath complicate the timeline, though the model includes an explicit COVID-19 interruption. Still, as one of the first studies to pit perception against millions of behavioral records, the work offers a sobering corrective to both hype and panic: at this institution, print, digital, and AI channels are coexisting, and the AI-replaces-libraries story remains, for now, unproven.

Subject of Research: How generative AI use relates to university students' continued use of library print and digital academic resources

Article Title: Generative artificial intelligence reshapes university students' sustainable use of academic information resources: mixed evidence from behavioral logs and perceptual survey data

Article References: Huang, Z., Song, X., Zhuang, J., & Xiao, S. (2026). Generative artificial intelligence reshapes university students' sustainable use of academic information resources: mixed evidence from behavioral logs and perceptual survey data. Frontiers in Psychology, 17, Article 2002741. https://doi.org/10.3389/fpsyg.2026.2002741

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.2002741

Keywords: generative AI, academic libraries, print borrowing, digital resources, structural equation modeling, interrupted time-series, cognitive offloading, self-regulated learning, substitutive use, auxiliary use, China University of Mining and Technology, information behavior