This pattern, sometimes called the trust paradox, does not appear to be a temporary problem to be solved by improvements in the technology. It is, rather, a structural feature of the situation. Artifi- cial intelligence systems are useful enough to be worth using, opaque enough to be worth distrusting, and sufficiently variable in perfor- mance that the same user moves reasonably between confidence and concern within a single session.
For library practice, the implication is that we cannot simply advise users either to embrace these tools or to avoid them. Both responses are inadequate. The right response is more nuanced and harder to teach: use the tools where they help, distrust them where the stakes are high (as in health situations), and develop habits of verification that match the situation.
Where artificial intelligence is appearing in library work
The reviews report artificial intelligence appearing in three broad areas of library and information work: in services to users, in research support, and in the back-office operations of the library itself.
Research library at the Richard Gilder Center for Science, Education and Innovation.
selecting which source to trust, weighing one source against another, recognising when sufficient information has been gathered, is now performed in significant part by the system itself, in ways that are invisible to the user. This is of consequence for libraries, since much of what we teach under the heading of information literacy concerns precisely that intermediary work. We teach users to evaluate sources, to triangulate, and to be cautious of the first item retrieved. When a student types a question into ChatGPT and receives a confident-sounding paragraph in reply, much of that work has either been done invisibly or omitted altogether. The reviews describe this as a shift from information liter- acy to artificial intelligence literacy. The new skills include the ability to construct an effective prompt (the new word for a search query), to recognise when a model is inventing or ‘hallucinating’, to push back when an answer seems too neat, and to triangulate the output against original sources. These are not the same skills as those of conventional infor- mation literacy, although they bear a family resemblance to them.
The trust paradox
If one finding emerged from nearly every review we read, it was this: users like artificial intelligence and worry about it, and they do both at the same time. Users praise these tools for being quick, conversational, easy to use, and capable of handling vague, half-formed questions of the kind they would never put to a search engine. They report enjoying the exchange, and they appreciate that follow-up questions can be asked in plain language. In studies of consumer behaviour, artificial intelligence performs well against the various technology acceptance factors: perfor- mance, ease of use, social influence, and the surrounding conditions that make the use of a tool feel ordinary. Yet the same users worry about hallucinations, about bias, about privacy, and about what becomes of the data they provide. They worry that the answers seem so polished that errors are difficult to detect. They worry, in academic settings, about plagiarism and integrity. In health-related settings, where the stakes are highest, the worries are sharpest of all.
Trust is not a matter of having confidence in artificial intelligence, or not having it. It is something that users negotiate repeatedly, depending on the task and what is at stake.
August-September 2026
Services to users. Library chatbots have become more common, particularly since the COVID pandemic obliged libraries to find ways of helping users they could not see in person. The reviews are largely positive about these implementations. Users appreciate the conversational design, the round-the-clock availability, and the ease of obtaining basic answers without the need to phrase a query in search-engine syntax. Where chatbots fall short is on emotional sensitivity, since they are poor at recognising when a user is upset or confused, and on accessibility, since some implementations are not friendly to users with disabilities. Libraries deploying chatbots need to consider both points carefully.
Research support. Artificial intelligence tools are being used by researchers, and by librarians supporting them, at every stage of the literature review process: in identifying relevant work, in screening for inclusion, in summarising findings, and in some measure in syn- thesis. The benefits are real, in the form of faster screening, broader discovery, and less drudgery, but so are the risks: fabricated refer- ences, missed sources of importance, and biased pre-screening. The current consensus, including from researchers who have tried these tools rigorously, is that artificial intelligence accelerates the work without replacing the human judgement at its centre. Researchers who attempt to remove the human from the loop tend to encounter difficulties. Libraries training researchers in artificial intelligence-as- sisted review work should teach this from the outset.
Back-office operations. The reviews touch on artificial intelli- gence in collection development, bibliometric analysis, cataloguing support, and similar tasks. The published literature is less developed here than on the user-facing side, but the direction is clear. Artificial intelligence tools can assist with routine, high-volume work, but the same caveats apply: human review remains necessary, and the cost of errors can be considerable if AI-generated metadata, summaries, or recommendations are placed into production without checking.
What the reviews say is missing
In addition to reporting what is known, every review we read also flagged what is not. The gaps in the literature are themselves inform- ative for libraries, since they indicate the questions that cannot yet be answered with confidence.
The first gap is theoretical. Most studies of artificial intelligence and information behaviour are using theories developed before arti- ficial intelligence existed. The classical models, including Kuhlthau’s information search process, Wilson’s problem-solving model, Bates’s berrypicking, and the technology acceptance model, were built for a world in which users searched databases and clicked on links. They were not built for a world in which the system generates the answer. Whether they continue to apply is an open question, and the reviews are divided. Some authors argue that the older models can be extended; others argue that the distinctive features of artificial intelligence, namely its generative capacity, its conversational interface, and its
INFORMATION PROFESSIONAL 17
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56
