An Argument for Libraries
Artificial Intelligence is changing libraries and services. While recent articles I have read focus on products that streamline librarian workflows or change how patrons interact with the collection, the more important question is how AI is affecting our relationship to knowledge. People tend to talk about AI as a tool—either for good or bad—something that does something for us. Chatbots answer virtual reference requests, and Claude summarizes and synthesizes articles for students. And that’s great, but it is also the problem. A professor friend of mine joked once that he was going to create a grading bot that would mark up student papers. He said, “Why not? They aren’t writing the papers, so why should I read them?” As amusing as the idea of one large language model grading another large language model’s work is, we’d have to admit we had lost the thread. But even if the actual scenario is never quite as absurd as the one I imagine, in becoming dependent on AI, we risk losing our willingness to learn, our ability to do, and ultimately the community that creates knowledge.
I admit, I would much rather play chess on my phone than write this essay, and if I asked Claude to write it for me, it would probably do a decent enough job. I bet with a solid revision after, you wouldn’t even know it. Once, pressed for time, I had AI look through three articles and two PowerPoint slides I had written and synthesize them into a new presentation. The product was decent, and after I had corrected some of the omissions and mistakes, it was acceptable. But most people are not asking Grok to summarize their own work, and they don’t know the subject well enough to know when a mistake has been made. Additionally, the next time I needed to put together a new presentation out of old material, I sat staring at my screen for 5 minutes unsure of how to proceed without simply having it done for me. My writing process is to start writing and write until I am done, editing and revising later. If I can’t even get started, it’s a big problem. Moreover, I imagine that the more one takes that shortcut, the harder it is to break the habit. Beyond that, the entire purpose of writing is to work out how and why you feel something and then share it with others who will hopefully engage in dialog with you. This is what Paulo Freire called communicative knowledge building—the notion that we learn through dialog with one another. That knowledge is, in many instances, the effort of a community participating in its own betterment. In that regard, having machines do it for us is ultimately to society’s detriment.
What is AI?
AI is often used as a shorthand to talk about a lot of things, but here I am referring to large language models that generate text/responses by predicting what the next word is most likely to be ( i.e., Grok, Claude, etc.). The easiest example is Markov chains, which were among the early ancestors of modern LLMs. Markov text generators were responsible for the predictive text options on your flip phone. But continuing to take the suggestions quickly turned into absurdity. The choices were grammatically correct but contextually unmoored because the program lacked memory. I used to play a predictive texting game with a friend. We would choose the center choice repeatedly and press send. The other would manually enter the first three words of the response and then let the Markov chains flow. After a while, we would be eating strawberries on a mountain serenading each other with lyrics from the musical Oklahoma. It was as bizarre to us as it was statistically defensible to our phones.
Modern LLMs have come a long way, having been trained on the equivalent of billions (if not trillions) of pages of text from all manner of sources, from The Brothers Karamazov to reddit posts. But AI isn’t sentient. It doesn’t understand what it has read or what it offers as an answer, it just assembles a pattern of language that is an amalgamation of the data it was trained on, and that pattern may answer your question with varying degrees of success. To be clear: there is no understanding, experience, or participation involved in the answer.
But what does this have to do with information literacy? Quite a bit, actually. AI is not just a tool that summarizes, synthesizes and answers questions. It shapes what information we see and how that information is presented. Increasingly, as its commercial use expands, AI becomes like the air we breathe, its presence so ubiquitous that we do not even notice it. But information literacy has always been about helping individuals connect with a given community of knowledge and participate in the creation of meaning. AI offers answers, but it does so in a way that encourages people to skip the most important parts of the process. Why learn a body of knowledge if you can just ask your computer to tell you what it says? Seeking knowledge from something that has neither experience nor understanding of the answer it gives you seems like a bad idea.
Source https://www.flickr.com/photos/152824664@N07/30212411048/ Author https://www.vpnsrus.com/
How Does AI Work?
I have spent enough time playing with AI (sorry trees and water and everyone else’s increased electricity bills) to recognize that AI is good at summarizing vast quantities of information, finding grammatical errors in documents, making safe (and bland) editorial recommendations, coding, and making me feel like I am saying the most insightful and important things ever. Likewise, a friend of mine has had an ongoing dialog with his computer for months about the nature of god. He describes the conversation as intellectually stimulating.
But where did the information come from, if not intelligence, artificial or otherwise? Back in the 1950s, Alan Turing proposed a computer that could be coded to learn and which would eventually create its own programming to work beyond the bounds of the original code. He also developed the Turing Test: If a human is carrying on two text-based conversations, one with a human and one with a machine, and cannot tell which is which, then the machine passes the test. The computer did not have to be intelligent, only mimic intelligence. LLMs have come a long way in this regard, with datasets consisting of billions of pages scraped from the internet, digitized books, articles, academic papers, encyclopedias, code repositories, and more. But before the machine spits out the world’s wisdom, the model has to go through quality control. Engineers devise mathematical models to score which datasets should be emphasized in results, for instance, preferencing Wikipedia over blog posts, and then a mathematical tool called a loss function is used to train the model to learn from its own mistakes and produce coherent writing instead of gibberish. Along the way, developers enter examples of desired answers and weight them so that the model has a paragon to aim for, and finally, the models are tuned through reinforcement learning from human feedback, where accurate (and usually confident and polite) answers are given a thumbs up and bad answers a thumbs down. This process likely played out in many different ways, as universities, corporations, militaries, and governments worldwide have all tried to harness machine learning.
AI and Authority
But for all this effort and input, is AI an authority we can trust? Stanford’s 2026 AI Index Report states that AI systems are becoming more powerful and are being deployed more rapidly than we can evaluate or govern. Stanford points out inaccuracy rates (i.e., hallucinations) of 22% to 94% across various agents and versions and that AI companies regularly publish performance reports but report on safety, fairness, and transparency with much less frequency. It would seem a natural outcome that documented AI-related incidents—that is, specific, real-world events in which AI systems have caused measurable harm, system failures, or data breaches–have increased. The irony being that because of their training, each LLM causing an incident did so with a smile and all the confidence in the world.
The Organisation for Economic Co-operation and Development (OECD) reports on an AI user in China who relied on Doubao to identify wild mushrooms and ended up with acute kidney failure. Moreover, 404 Media and the Guardian have reported on AI-generated books on mushroom foraging that have given unorthodox (and potentially deadly) advice.
The problem isn’t just mushrooms, though. A recent 11th Circuit States Courts of Appeals decision recently found need to include the statement
But it isn’t all bad. In fact, some of it is pretty amazing. AI’s ability to model complex structures and assimilate enormous volumes of data have led to breakthroughs in Alzheimer’s research, tuberculosis therapies, cancer treatment, to name a few. The same technology responsible for dangerous misinformation is simultaneously capable of remarkable discoveries—all with equal confidence. So how do we ensure that we keep the good parts of AI and get rid of the bad?
Regulating AI
AI regulation and oversight significantly lag development and deployment. It has seemed a bit like the wild west for the past few years, but there are indications that oversight is finally catching up. Recently, a German court found that Google is liable for false statements generated by its AI products and some international and state governments are requiring chatbots to disclose that they are, in fact, not human.
Laws that make companies responsible for bad information created by their products would seem to be common sense. The need for oversight, according to Karni Chagal-Feferkorn, an assistant professor who researches AI law, ethics and policy in USF’s Bellini College of Artificial Intelligence, is urgent. To underscore this claim, she references AI companions that have encouraged suicide and homicide, among other bad pieces of advice. Finally, she brings it back to something equally urgent that we do not necessarily talk about:
“Children’s overreliance on AI could potentially lead to loneliness and depression, along with diminished social skills and problem-solving abilities. Another concern is how much information is collected about users and how it might be used later in ways that don’t benefit them.”
These concerns are squarely in the wheelhouse of information literacy.
AI and Information Literacy
If some of the concerns about AI are inaccuracy, loneliness, loss of critical thinking, and erosion of privacy, the library would seem to be the old, reliable antidote. Libraries have long positioned themselves as nexus points for community, be it through events, programming, or public meeting spaces. Just as surely, the ACRL embeds critical thinking into every aspect of the Framework and the Library Bill of Rights is unequivocal in its advocacy of patron privacy. But in the end, all of these things require a user to know what services a library offers, to put in the effort to learn, and to participate in social situations. Frankly, for a lot of people, it is easier to talk to a computer, accuracy be damned.
There are, however, some practical steps we can encourage. We start by acknowledging that the outcomes of AI use are widely variant—either eating poisonous mushrooms, rendering career- and credibility-destroying inaccuracies OR ushering in medical breakthroughs. We must try to teach patrons how to know when to trust it and when not to.
It is a bit of a reach to expect our patrons to know whether AI is right or wrong about protein pathways that trigger Alzheimer’s, but because AI is so frequently wrong, again, giving answers by statistically predicting the next word instead of relying on a comprehensive understanding of a body of literature, it is a good habit to check the sources AI agents cite and then to evaluate those sites, instead of the answer provided. AI doesn’t always have the best grasp of accuracy, so looking at the input behind the answer is necessary. For this task, librarians have long recommended evaluation rubrics, whether CRAAP, SEER, PROVEN, or one of the many others. By evaluating the inputs for objectivity, accuracy, and timeliness, users can hopefully catch discrepancies between input and output, as well as learn some context and nuance along the way.
Information literacy–even before the Framework—was not just about evaluating the information you encounter in the world. It also stressed the importance of being ethical producers of information. In a world wrestling with AI, this is even more important. Patrons need to know about inaccuracy rates and the dangers of putting their name on information that has not been verified. Likewise, creatives need to know that many publishers will not accept material that has been generated by AI, and that such materials sometimes cannot be copyrighted. AI also makes it easy to create images out of whole cloth and pass them off as truth, so we need to encourage ethical use of these systems. But most importantly, we must build inclusive and welcoming spaces where people can be part of a community, increase programming on critical thinking about information sources, and continue our vigilance about patron privacy.
A proponent of AI will likely read this essay and laugh. There is no way libraries can keep up with AI—not in convenience, speed, or scale.
True.
But libraries were never primarily about convenience. They are about participation. They help people learn, practice, question, create, and engage with others. They are among the few remaining institutions dedicated not simply to providing information, but to helping communities make sense of it.
In the end, my concern about AI is not that it will become more intelligent than people. It is that we may become less willing to learn, less capable of doing, and less connected to the communities through which knowledge is created. What libraries can provide is human connection, human expertise, and human experience. AI should be used to supplement those things, but it cannot be used to replace them.
About the Author
Todd Heldt is a librarian, educator, and writer whose work focuses on information literacy, critical thinking, artificial intelligence, and open educational resources. He is the creator of LIS101.com. His personal website is toddheldt.info.