© Copyright 2026 Mastersinlibraryscience.org.Â
What a new study's circulation and platform logs mean for library assessment
Students now ask an AI chatbot before opening the library catalog, yet many academic libraries still assess use through print circulation and database logins. A Frontiers in Psychology study treats those logs as evidence of how generative AI reshapes students' sustainable use of academic information resources1.
The study pairs behavioral logs with survey data and applies interrupted time-series segmented regression to monthly loan volume. That makes AI-feature activity on digital platforms a metric to track alongside circulation, not a side conversation.
The implication: library assessment and collection management can no longer separate AI behavior from baseline borrowing patterns.
Published in Frontiers in Psychology on 26 October 2026 as Volume 17, Article 2002741, the study appears in the Educational Psychology section and in the Research Topic "Generative AI in Education: Cognitive Mechanisms, Learning Outcomes, and Evidence-Based Practice - Volume II."1 It asks how generative AI reshapes university students' sustainable use of academic information resources, pairing a survey of university students with institutional behavioral logs, including nine years of library loan data and platform operations.1
The paper's available limits include perceptual associations not matching observed behavior, auxiliary use overlapping with perceived convenience, and no clear January 2023 loan breakpoint; it does not report a sample size or institution count in the available sections.1
Read the full study in Frontiers in Psychology.
The central tension in library assessment is balancing what students say they do with what operational logs actually record. The Frontiers in Psychology study addresses this by pairing two kinds of evidence: behavioral logs from print circulation and a digital resource platform, plus perceptual survey responses from students.
The print-book circulation table reports annual indicators, giving libraries a concrete view of loan volume over time. A separate platform table tracks annual trends in traditional operations and AI-feature operations, making it possible to see whether AI-assisted activity is rising alongside, or instead of, conventional digital use. These two tables together form the study's behavioral core: loan activity and platform operations are measured as recurring, comparable indicators, not one-off snapshots.
The survey side uses latent-variable checks, including HTMT, Fornell-Larcker, and model fit statistics. These checks make the perceptual findings more credible, and they offer a replicable template for LIS researchers who want to combine logs with student reports. Researchers can use this validity reporting to show that survey constructs measure what they claim before connecting them to log-based outcomes.
The assessment implication is clear: AI-feature activity belongs next to print circulation and traditional platform metrics in collection-management reports. But the evidence is observational, so librarians should describe shifts in use rather than claiming AI caused them. This keeps the conversation about sustainable information use grounded in evidence that is visible in dashboards and annual reports, not just in student anecdotes.
How can a library tell whether a new AI feature actually shifted monthly borrowing rather than just riding an existing trend?
Interrupted time-series (ITS) segmented regression compares the pre-intervention trend with the post-intervention pattern.1 It estimates two changes at the AI-feature launch date: the immediate jump or drop (level change) and the difference in month-to-month trajectory afterward (slope change). Suppose loans rise by 1,500 items at the switch and then add 300 more items per month than before.1 That pattern, not a single snapshot, signals an association.
Table 9 of the Frontiers in Psychology study reports segmented regression results for monthly loan volume, so libraries can treat it as a worked example. Export monthly loans from your ILS or platform log, mark the launch date, and code time as months since the intervention. Ordinary linear regression suits continuous outcomes; Poisson models fit count data.1
Students face a first decision: blindly trust an AI answer or treat every cited source as unverified until it appears in the catalog. Accuracy research is task-specific, not a single reliability score.
Published benchmarks vary widely. In a 2026 medical literature search, hallucinated citations ranged from 0% for OpenEvidence and 3.0% for Perplexity to 27% for Claude, 42% for Copilot, 63% for Gemini, and 70% overall for ChatGPT (81% for unique references). A 2025 eight-chatbot comparison found 39.8% of references wrong or fabricated, with Copilot at 100%, Perplexity at 72%, Claude at 64%, and Grok and DeepSeek at 0% for tested prompts. Another 2026 test reported 55.0% of references had at least one error, 28.3% completely fabricated. Across 13 large language models, hallucination rates span 14% to 95%. These are version- and prompt-dependent, not general reliability. No robust independent citation-fabrication rate has been published for library-vendor AI research assistants, so treat vendor benchmarks cautiously.
For undergraduates, focus on basic source existence and type. For graduate students, emphasize citation tracing, discipline-specific databases, and prompt design. Published outcome data for library AI literacy workshops remains limited: a 2026 guide lists six repeatable outcomes but reports no sample sizes, pre/post scores, or effect sizes. Be especially cautious about claiming measured gains.
How much energy does an AI answer actually use, and does that footprint change how academic librarians should teach information seeking? The study on generative AI and students' information use does not measure carbon, so this section draws on outside energy research.
In library practice, sustainable use can mean efficient, durable use of institutional information resources, or the environmental footprint of AI tools. Librarians should keep both visible in instruction and procurement.
The often-cited 0.3 watt-hours per conventional web search dates to 2009 and should not be treated as today's baseline.1 Recent typical generative text-query estimates cluster around 0.24 to 0.3 Wh, close to that older figure rather than dramatically higher.1 Google reports 0.24 Wh and 0.03 grams CO2e per Gemini Apps text-only prompt, while independent estimates put GPT-4 at 0.43 Wh and GPT-4o at 0.42 to 1.79 Wh.2 Response length matters: one estimate is 0.3 Wh for under 100 words but 3.5 Wh for over 7,000 words.3 A 2026 study of frontier models found a median 0.31 Wh for typical queries, while test-time reasoning reached a median 3.91 Wh.4
The International Energy Agency adds context: simple AI text queries typically use less electricity than a television over the same period, and replacing all conventional searches with simple text queries would stay below 1 percent of data center consumption, though reasoning, agentic, and video-generation tools can use hundreds or thousands times more.5 Estimates vary with model, tokens, prompt length, hardware, utilization, accounting boundary, and electricity mix, so they are contested and not interchangeable.1
Teach students to choose the right tool for the task, avoid repeated prompts when the first answer is close, and use the catalog or a database for known-item searches. That reduces AI demand and models a more deliberate definition of sustainable information use.
A procurement checklist reframes generative AI governance as a set of vendor questions rather than a ban or allow debate. The criteria below draw on university library guidance and publisher terms, covering publisher permission, closed environments, training use, fair use, data protections, and negotiated rights.
| Criterion | Question to Ask the Vendor | Framework or Source |
|---|---|---|
| Publisher permission and closed environment | Does this tool require publisher permission and operate only in a closed, enterprise-grade environment limited to current university users, while excluding public AI tools? | Artificial Intelligence and Licensed Content: Home (University of Chicago Library) |
| Licensed content terms for training | Does the vendor permit only limited portions of licensed content in closed AI tools for noncommercial research or academic purposes, and does that permission exclude training, fine-tuning, testing, or evaluation? | Wiley licensed content |
| Uploading to third-party platforms | Is the library-licensed e-resource permitted to be uploaded to a third-party platform, including a personal generative AI account, or is that generally not permitted? | Guidelines for using library-licensed content (University of Toronto Libraries) |
| Text and data mining vs generative AI rights | Does the license allow text and data mining but require separate checking of generative AI use rights? | Use of library resources with AI (University of Waterloo Libraries) |
| Publisher terms and electronic resource licenses | Does the planned use comply with publisher terms, license agreements, copyright law, and university policies, especially where electronic-resource licenses restrict uploading materials into third-party AI platforms? | Statement on Electronic Resources Use (Carnegie Mellon University Libraries) |
| Noncommercial use and training data | Does the tool permit noncommercial research, private study, or teaching with ITS-supported AI tools that do not copy or use the material as training data, while noting negotiated licenses may override fair-use rights? | Copyright and Licensed Materials: Using Generative AI in Research (University of Southern California Libraries) |
| License terms vs fair use | Does uploading licensed articles, book chapters, datasets, images, or other content to a third-party AI tool violate license terms even if a separate fair-use argument exists? | Copyright and AI (Louisiana State University Libraries) |
| Data protections and approved tools | Does the tool provide protections for users and their data, and has it been vetted and approved by the institution before sensitive data is used, especially where publisher agreements may prohibit text and data mining? | AI, Copyright, and Research: A Guide (Columbia University Libraries) |
| Contractual restrictions and negotiated rights | Does the library's negotiated license include text and data mining and AI training rights, or do contractual restrictions limit AI training with licensed material? | Generative Artificial Intelligence for Teaching, Research and Learning (University of California, Davis Library) |
In a 2023 College & Research Libraries survey of 760 academic library employees, 62.91% said they did not feel adequately prepared to use generative AI professionally. Only 3.68% were classified as high AI literacy1, and 69.76% reported low confidence troubleshooting AI tools. Ethical evaluation fared somewhat better: 29.50% reported low confidence evaluating ethical implications, while 74.34% agreed ethical and privacy concerns were urgent.
Start with basic tool literacy: hands-on practice with common chatbots, prompting, and recognizing outputs. Next, move to verification and instruction skills, teaching staff to help students fact-check AI-generated citations and summaries against library databases. Finally, add governance and assessment: procurement checklists, privacy review, and interpreting usage logs. The 2024 AI literacy framework's seven competencies, covering capabilities, use cases, ethics, privacy, and stakeholder impacts, map neatly onto these rungs.
The survey shows 43.99% rated AI training for librarians as extremely important in the next twelve months. Preferred formats were online courses or webinars (26.02%) and self-paced modules (22.44%). Staff who can read circulation and platform logs are better positioned to evaluate whether AI features change student use, tying this training back to the interrupted time-series work discussed earlier.
