Artificial intelligence (AI) and information seeking: A comparative exploration of AI chatbots, search engines, and library resources as information sources among university students

Synthesis: Lund and colleagues survey 236 US university students about where they start an academic search and what they think of what they find, comparing AI chatbots, search engines and library resources. Search engines remain dominant — Google 45.5% of preferred starting points against the library homepage at 27.9% and an AI chatbot at 26.6% — and students rate search engines more relevant and more satisfying for nearly every information type. But AI is not marginal: 26.6% start there, 9.8% use it daily for academic work, and the demographic split is sharp. Younger students and international students use AI far more, and international students report using it as a replacement for library resources rather than a supplement.

Key Findings

1. Preferred place to begin an academic search: Google 45.5%, library homepage 27.9%, AI chatbot or LLM 26.6%.
2. Reported uses in the past month average 18.71 for Google, 14.73 for library resources and 12.95 for generative AI.
3. Generative AI use distribution: never 15.3%, less than monthly 43.0%, monthly 24.3%, weekly 7.7%, daily 9.8%. Library databases: never 1.7%, daily 3.8%.
4. Age predicts the starting point (χ² = 38.87, p < 0.01): among 18–25-year-olds, Google 52%, AI tools 32% and the library homepage 16%; among students 31 and older, the library homepage is 54%, Google 23% and AI 23%.
5. International students substitute AI for library resources: mean monthly AI uses 16.76 versus 7.86 for domestic students (F = 28.09, p < 0.01), while library use is 12.51 versus 17.78 (F = 13.48, p < 0.01). Google use is nearly identical (18.99 versus 19.08).
6. About 28.1% have never used generative AI alone for academic tasks, and 24.4% have never used library databases or search engines first and then AI; 58% of domestic students have never used AI alone against 8% of international students (F = 95.33, p < 0.01).
7. Confidence that an answer is complete: 58.7% are somewhat or very confident in generative AI, against 48.9% who are very confident in library databases; yet 19.6% have no confidence at all in AI's completeness while only 15.3% never use it.
8. AI loses to search engines on perceived relevance, likelihood of use and satisfaction for job-related, coursework, exam, weather, news and storytelling information; the two tie only on instructions for performing a task.
9. Frequency of AI use tracks satisfaction with AI (r = 0.31 to 0.57 across information types, all p < 0.01) but not satisfaction with search engines, whose means are near-identical between frequent and infrequent AI users.
10. Gender shows a bimodal pattern (χ² = 98.03, p < 0.01): 13.2% of men never use AI against 2.2% of women, while 11.4% of men use it daily against 3.3% of women.

What the substitution pattern means

The international-student finding is the one a library cannot ignore. Domestic and international students use Google at effectively the same rate; the difference sits entirely between the library and the chatbot. That is not a preference for novelty — it is a substitution, and it is large enough to explain why a library's instruction program can look healthy in aggregate while a substantial part of its audience has stopped arriving through its door.

The satisfaction finding complicates the usual story about over-trust. Students who use AI most are the most satisfied with it, but the direction of that relationship is unmeasured: heavy use may produce satisfaction, or satisfied users may use more. What the data do show is that AI is rated below search engines on relevance for almost every information type, including coursework. Students are not treating the chatbot as a better search engine; they are treating it as a different kind of source, and one they are less sure about.

A small, skewed sample with a clean instrument

The sample is 236 students recruited largely by convenience from information-science programs, 84% graduate and 81% in natural and computing sciences, collected in late summer 2025. Those proportions bound every percentage above: this is a technically oriented, mostly graduate population, not a campus cross-section.

Within that limit the instrument is unusually transparent. It adapts items from a national study of academic researchers' information seeking and from research on voice assistants and search-to-AI switching, is reproduced in full, and reports the frequency, relevance, use and satisfaction measures separately. That makes the paper more useful than its sample size suggests: the question set is reusable, and the comparative design — three sources, seven information types — is a template a library could run on its own population.

What this means for practice

• Segment instruction by what students already do. Age and international status predict the starting point more strongly than anything else measured; a single session for all students addresses a population that does not exist.
• Treat international students as a design target, not a diversity footnote. The substitution is largest there and the support need — language, source evaluation, local academic norms — is specific.
• Ask about completeness, not just accuracy. A fifth of users have no confidence in AI's completeness yet keep using it; instruction that teaches verification addresses a doubt students already hold.
• Instrument the door, not only the desk. If a quarter of academic searches start outside the library, usage counts from the library's own systems will understate the change.

Limitations

• The sample is 236 students, recruited largely by convenience from information-science programs; 84% are graduate students and 81% study natural and computing sciences, so the results describe a technical, graduate-heavy population.
• All measures are self-reported frequencies, relevance ratings and satisfaction on Likert scales; there are no behavioral logs or task outcomes.
• The survey was collected in late summer 2025 and describes a moving target as tools and norms change.
• The analysis is descriptive and correlational; the study states no causal claim about why students choose one source over another.
• The reported per-group Google means (19.08 domestic, 18.99 international) do not reconcile with the overall mean of 18.71 at the stated 43/57 split, an internal inconsistency in the paper.

Connected Concepts

• Information and Media Literacy
• Libraries
• Librarians
• Generative AI
• Conversational AI
• Trust Calibration

Citation

Lund, B. D., Teel, Z. A., Mohammed, Y., Jagathpally, A., & Wang, T. (2026). Artificial intelligence (AI) and information seeking: A comparative exploration of AI chatbots, search engines, and library resources as information sources among university students. Journal of Librarianship and Information Science, 1–14.