Did generative AI replace university library borrowing?
Not measurably at one Chinese university. A nine-year, 3.2-million-transaction record found no significant change in print borrowing after January 2023, while a 368-student survey showed only a weak substitution signal.
The short version
- What happened
- The study combined 368 questionnaires with 3.2 million print-library transactions from January 2017 to December 2025 and digital-platform statistics through the first half of 2026.
- Who it affects
- Students may need different tools for deep reading, verified retrieval and rapid synthesis; a single 'AI versus library' policy would miss those task differences.
- What remains unproven
- One university in China
Keep in mind
One university in China
Cross-sectional survey of 368 students
Survey responses could not be linked to individual borrowing records
Several constructs used single questionnaire items
The interrupted series cannot isolate every post-2023 institutional change
No learning, comprehension or academic-integrity outcome
Source analysis · 1 linked source. Primary analysed source date: 26 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The long record did not show an AI break
Generative AI had not measurably accelerated the decline in print borrowing at the China University of Mining and Technology by the end of 2025. Researchers analysed 107 months of circulation data—about 3.2 million borrowing, return and renewal records—from January 2017 to December 2025. February 2020 was unavailable, and the model separately represented the pandemic shock, seasonality and January 2023 as the prespecified GenAI breakpoint.
Print loans were already falling by about 536 books a month before the pandemic. The estimated January 2023 level change was 1,167 additional books and the later slope change was 160 fewer books a month, but neither result was statistically significant. Alternative breakpoints in March 2023, September 2023 and January 2025 gave the same null pattern. A null result does not prove no student changed behaviour; it means the aggregate record did not detect a new institution-wide break.[1]
Survey answers and platform use pointed to coexistence
A separate 2026 questionnaire of 368 enrolled students distinguished auxiliary use—searching, summarising and comparing—from substitutive use, such as generating complete assignment passages. Substitutive use had a weak negative association with intention to keep using print resources (standardised coefficient −0.100; effect size f² 0.104). Auxiliary use was positively associated with continued digital-resource use, although those two survey constructs were not statistically distinct enough to support a clean mechanism.
Digital-platform totals also resisted a simple replacement story. Traditional operations rose from 280 million in 2021 to 317 million in 2025. Embedded AI features recorded 326,000 operations in 2025 and 766,000 in the first half of 2026, but represented only 0.10% and 0.53% of platform activity. That is early uptake layered onto existing services, not evidence that AI displaced them.[1]
The distinction that matters
Three evidence streams, three different questions
The study triangulates behaviour and perception without claiming they are the same dataset.
Print logs
3.2m transactions over 107 months
No significant post-January 2023 break
Student survey
368 cross-sectional responses
Weak print-substitution association
Digital platform
Operations from 2021 to mid-2026
AI grew from a very small base alongside traditional use
Libraries should test tasks, not assume extinction
For librarians and teachers, the result supports designing around different tasks: print for sustained reading, licensed databases for sourced retrieval and AI for navigation or synthesis with verification. It does not justify cutting collections, nor does it establish that AI improves learning. The study measured intentions, transactions and platform operations—not comprehension, grades, source quality or academic integrity.
The survey and logs could not be linked person by person, several survey concepts used single questions, and all data came from one research-intensive Chinese university. Funding was reported from the Chinese Academy of Agricultural Sciences' Agricultural Information Institute. Replication across institutions, linked privacy-preserving usage records and measured learning outcomes would show whether coexistence persists as embedded AI becomes a larger share of academic-resource use.[1]