What AI actually is

When a librarian types a question into ChatGPT or Claude, the model does not ordinarily search a database or retrieve stored facts. The exception is a tool connected to search or to a document collection, as discussed later in this module. Otherwise, it predicts what text should come next, based on statistical patterns learned from billions of documents during training. For example, if a librarian asks about the history of interlibrary loan, the model draws on patterns from everything written about that topic in its training data, including journal articles, library blogs, textbooks, and websites. It then constructs a response by predicting what a plausible and informative answer would look like. The model is an extraordinarily sophisticated pattern matcher, but it does not retrieve stored facts the way a database retrieves records.

To understand this at a slightly deeper level, it helps to know how a language model processes text. Models do not read words; they read tokens, which are fragments of text averaging roughly three-quarters of a word. For example, the word "cataloging" might be a single token, while "interlibrary" might be processed as two. The model converts every token in a prompt into numbers and passes those numbers through layers of mathematical operations called a transformer architecture. It then generates a response token by token, choosing each one by the statistical probability of what should follow everything before it. This process produces remarkably fluent text because fluency, grammatical coherence, and topic relevance are precisely what the model learned to maximize during training.

This distinction matters considerably for library practice. A database returns records; a search engine returns links; a language model generates text that sounds plausible based on what it has learned. Plausible is not the same as accurate, and the difference shapes how librarians should integrate these tools into professional work. Understanding the mechanism, prediction rather than retrieval, is the single most important conceptual foundation for using generative AI well.

This module covers what Leo S. Lo's AI Literacy for All: A Universal Framework calls Technical Knowledge, the first of its five components. It focuses on that component's first two levels: defining AI and recognizing its common applications, distinguishing AI myths from realities, explaining how machines learn through pattern recognition, and recognizing the role of data and the limits of the technology (Lo, 2025). Lo describes this component as demystification rather than specialization, and the distinction is useful here. A librarian does not need the mathematics of a neural network to reason accurately about what a language model can and cannot do. That working level of understanding is exactly what this module aims for.

Before a model can predict text fluently, it must learn from an enormous quantity of text. This process, called pretraining, exposes the model to hundreds of billions of words from books, websites, academic papers, code repositories, and other sources. The model's internal parameters are adjusted until it becomes very good at predicting what comes next in any given sequence. For example, the models behind ChatGPT, Claude, and Gemini are trained on large web crawls, books, Wikipedia, and code from public repositories. That text is collected up to a training cutoff date, which typically falls several months to a year before the model's public release. Each developer weights these sources differently, and none discloses the full mix.

Several implications of this training process matter directly for library practice. First, training data has a cutoff date, the point after which no new information was incorporated. For example, a model trained through early 2024 has no knowledge of research published, events that occurred, or databases that changed after that date. Any question requiring current information will therefore produce outdated or fabricated responses unless the user supplies that information in the prompt. Training data also reflects whatever was written on the internet, including the biases, assumptions, and errors in that broader written record.

Second, training data for major AI models included a substantial quantity of copyrighted material, specifically books, articles, and other published works, collected without explicit permission from rights holders. This has produced significant legal challenges. The New York Times Company v. Microsoft Corporation, filed in December 2023 against Microsoft and OpenAI, is one prominent example, and multiple authors have brought class-action suits over training data use. For librarians in institutions with active intellectual property policies, this is not merely background. It bears on professional decisions about which AI tools to recommend and how to discuss AI with faculty and students doing original research.

After pretraining, models undergo fine-tuning and reinforcement learning from human feedback. In this process, human reviewers rate model outputs, and those ratings are used to adjust the model toward more helpful, accurate, and seemingly safe responses. This refinement is why modern models tend to decline harmful requests and acknowledge uncertainty, but it does not guarantee factual accuracy. It does make them better at presenting information in ways that seem trustworthy. Paradoxically, that can increase the risk of accepting plausible errors without verification.

The training and prediction processes described above are not free. A complete picture of what AI is has to include what it costs the physical world to run. Training a single large model consumes an enormous quantity of electricity, and every later query draws additional energy from a data center that must be powered and cooled around the clock. For example, researchers estimating the water footprint of AI have found that a conversation of roughly ten to fifty responses with a model of the GPT-3 generation can consume about half a liter of fresh water, depending on where and when it runs (Li et al., 2023). The water goes to cooling, which keeps data center servers from overheating. These costs are invisible at the keyboard, which is exactly why they are easy to overlook. A librarian teaching AI literacy is well positioned to make them visible.

Three costs are worth naming specifically. The first is energy and the carbon that accompanies it. The International Energy Agency projects that electricity demand from data centers, driven substantially by AI, is rising steeply, and the carbon consequence depends heavily on whether that electricity comes from clean or fossil sources. The second is water, the cooling cost just described, which is a particular concern where data centers sit in already water-stressed regions. The third is electronic waste. The specialized hardware that runs AI has a short useful life and is replaced rapidly, and researchers have projected that generative AI could add substantially to the world's e-waste stream over the coming years (Wang et al., 2024). These costs are real, cumulative, and unevenly distributed. They fall hardest on the communities where the infrastructure is built rather than on the people sending the prompts.

A practical implication follows, and it is one librarians can act on. The largest general-purpose models are the most expensive to run in every one of these dimensions, and many library tasks do not require them. For example, summarizing a document a librarian pastes in, or drafting a routine email, can often be done well by a smaller, task-specific model that consumes a fraction of the energy of a frontier model. Choosing the smaller tool when it is sufficient is both an environmental decision and a professional one. It reflects what the American Library Association's AI guidance calls a "sufficiency mindset," which it recommends libraries apply to AI use "to mitigate environmental harm." Matching the tool to the task, rather than reaching for the largest model by default, is a judgment that remains with the librarian who understands both the work and its cost.

Hallucination, the term used when AI generates false information with apparent confidence, is not a simple bug awaiting a patch. It is a structural consequence of how language models work. Newer models hallucinate less often on many tasks, but none has eliminated it. For example, if a librarian asks Claude to provide three peer-reviewed sources supporting a particular argument about digital preservation, the model may generate three citations that do not exist. They can still look entirely plausible, with realistic author names, recognizable journal titles, and credible publication years. The model was not attempting to deceive; it was doing what it always does: predicting what a plausible response would look like. Statistically, a convincing citation is simply a plausible set of tokens given a question about sources. These fabrications are particularly dangerous in library contexts precisely because they look authoritative.

The model does not know what it does not know. A human expert can say "I am not certain about that; let me check." A language model has no mechanism for distinguishing information it learned reliably from text it is generating probabilistically, because both come from the same prediction process. The model may add disclaimers such as "I should note that I am not certain about this," but these disclaimers are also generated probabilistically. They appear when the model has learned that disclaimers fit certain contexts, not because it has verified its output.

To protect patrons and the professional credibility libraries depend on, librarians must verify any factual claims, statistics, and citations produced by AI before using them in professional work. There is no exception to this rule, however confident or well-sourced the response appears. It also helps to tell patrons directly: AI is not a research tool; it is a drafting and thinking tool, and anything it tells you about facts requires verification with a real source. This framing helps patrons see why the librarian's role in source evaluation remains indispensable, rather than diminished, by the presence of AI.

Dakan and Feller offer an analogy useful for patron instruction: hallucination is like a friend who recounts a story with absolute confidence while getting the details entirely wrong (Dakan & Feller, AI Fluency: Framework & Foundations, Anthropic Academy, 2025). The framing works because it removes any implication of deception. A hallucinated answer is indistinguishable in tone from an accurate one. That is exactly why verification is a non-negotiable professional step rather than an optional precaution, however authoritative a response appears.

Generative AI responses are probabilistic. Each time a question is posed, the model samples from a distribution of probable next tokens. It does not deterministically select the single "best" response; it draws from a range of plausible options weighted by probability. For example, a reference librarian who asks Claude to draft a database instruction email on Monday may receive different emphasis, examples, and phrasing than the same prompt produces on Friday. This variability can occur even within a single conversation when a prompt is resubmitted unchanged.

This unpredictability has direct implications for library practice. You cannot treat generative AI output as a stable, citable source, nor assume that because AI produced a particular answer once, it will produce the same answer again. The same prompt submitted to different AI tools will also produce different outputs. Each model has different training data, architectural choices, and fine-tuning, and all of these affect how it weights plausible responses.

It helps to think of a language model as a well-read colleague who may explain the same topic differently each time you ask. Their responses must therefore be evaluated on their own terms rather than assumed to be consistent. This framing also explains to patrons why they cannot simply accept AI output as authoritative. AI is not always wrong. The problem is that the same question does not reliably produce the same answer, and there is no way to know in advance which response is more accurate without independent verification.

AI performs certain tasks reliably and others poorly, and understanding the difference is essential for integrating these tools into library workflows. Assuming that AI is either uniformly capable or uniformly unreliable produces professional errors in both directions.

Whether AI suits a given task therefore depends on the task itself, the consequences of an error, and whether the output can be independently verified. These patterns make the professional role of the librarian indispensable. Generative AI functions as a drafting and thinking partner, not as a reference source, and this distinction should inform every decision about how and when to use it.

It is useful to distinguish three broad categories of AI. They function differently and raise different professional considerations for library practice.

Generative AI creates new content: text, images, audio, video, and code. ChatGPT, Claude, Gemini, and Perplexity all fall into this category. This is what most librarians are currently experimenting with, and it is the primary focus of Levels 1 and 2 of this curriculum. For example, when a librarian uses ChatGPT to draft a LibGuide introduction or Claude to summarize a lengthy accreditation report, that is generative AI. These tools are powerful for drafting, editing, and ideation. They require the verification practices discussed above because they generate rather than retrieve.

Predictive AI does not create new content; it makes recommendations and classifications based on patterns. For example, the "you may also like" systems embedded in discovery layers and integrated library systems are predictive AI that libraries have used for years, often without calling it AI at all. Spam filters, recommendation engines in library discovery systems, and automated metadata enrichment tools all fall into this category. These systems are familiar, if not always recognized as AI. They raise different questions than generative AI does, chiefly about algorithmic bias, data quality, and the transparency of automated decisions.

Agentic AI takes autonomous actions. It does not simply respond to prompts; it executes multistep tasks with limited human intervention, such as browsing the web, writing and running code, sending emails, and interacting with other software on the user's behalf. This is a newer and rapidly evolving category, covered in depth in Module 15. Understanding how agentic AI differs from generative AI matters increasingly for librarians involved in workflow automation, systems integration, and institutional AI policy. Agentic systems raise considerably higher stakes for oversight and accountability than tools that only generate text.

Two terms appear constantly in current discussions of library AI. ACRL's AI Competencies for Academic Library Workers (October 2025) names both explicitly within its Knowledge and Understanding competency category, which makes them foundational vocabulary for this entire curriculum.

The first term is retrieval-augmented generation, commonly abbreviated RAG. In a RAG system, the model does not generate a response purely from what it learned during training, the prediction process described earlier in this module. To ground its output in something more reliable than training data alone, it first retrieves relevant documents or passages from a defined knowledge base. It then generates its response from those retrieved sources. This grounding is what distinguishes RAG from pure prediction: the output is anchored to documents that actually exist, not only to patterns learned during training. For example, Ex Libris's Primo Research Assistant works this way. It retrieves candidate records from its discovery index before generating a response, which is why its answers come with citations to records a patron can actually open. Retrieval does not eliminate error, however. The system may select irrelevant sources, misread the evidence it retrieves, or draw conclusions the documents do not support. Retrieved sources and their citations therefore still require the same evaluation as any other output. Module 03 covers several library-specific tools built on this principle.

The second term is agentic AI, introduced just above. ACRL defines agentic systems as those that "set goals, plan tasks, and act with minimal guidance," a meaningful step beyond a chatbot that only responds to one prompt at a time. Module 15 explores agentic AI in depth; for now, the working distinction worth holding onto is this one: a chatbot answers, an agent acts.

Dakan and Feller's AI Fluency: Framework & Foundations course offers a practical distinction: it separates AI engagement into three modes (Dakan & Feller, Anthropic Academy, 2025). The first is automation: the task is defined, and the AI executes it. For example, a librarian might ask the AI to draft a patron-facing email about new database access, summarize a long policy document for staff, or generate subject heading options for a record under review. These tasks work well when the outcome is clearly defined and the professional knows exactly what output is needed. The second mode is augmentation, in which the librarian works through a problem together with the AI rather than handing it off. For example, when developing a new information literacy workshop, the AI can serve as a thinking partner. It can test different framings, push back on assumptions, or generate contrasting approaches before a direction is chosen. The AI does not do the work; it makes the work better. The third mode is agency, in which an AI system acts within parameters established in advance. Examples include a workflow that routes incoming research requests, or a chatbot that handles common after-hours reference questions before a librarian reviews them. These configurations require the most careful setup and the clearest understanding of what the AI may and may not do independently. Recognizing which mode a task calls for is itself a professional judgment that no AI can make on the librarian's behalf.

Librarians are among the professionals best positioned to understand why AI is not a search engine, because they already understand what search engines are and how they differ from databases. For example, when a student asks "Can I just use Google?" about library databases, librarians explain clearly why, for certain research purposes, they cannot. The same professional clarity is valuable when explaining to patrons and colleagues why AI is not Google, either.

The fundamental difference is as follows. A database stores records and retrieves them in response to queries. When a librarian searches JSTOR for articles on interlibrary loan, JSTOR returns actual articles that exist, were published, and are accessible in full text. The database is an index pointing to real documents. A search engine crawls and indexes the web and returns links to pages that exist at the time of crawling. The results are not always accurate, current, or authoritative, but they point to real, externally verifiable pages. An AI language model generates text by predicting probable continuations. It does not retrieve records or links; it produces new text. There is no document behind the response. This distinction is not intuitive for users who have spent their entire lives seeking information through retrieval systems.

One framing consistently helps patrons and colleagues grasp this distinction. A database is the library's stacks, organized and retrievable. A search engine is a map of the stacks, helpful for navigation. An AI language model is a knowledgeable colleague who has read a great deal and can discuss it, but who may misremember, conflate, and occasionally confabulate, and whose memory ends at a specific date. Such a colleague is valuable for certain purposes and wholly unreliable for others. Knowing which is which is the essential professional judgment.

This comparison also clarifies the role of AI in library instruction. Librarians who teach information literacy should not simply add AI to existing instruction about databases and search engines as though it were another tool in the same category. A language model is a fundamentally different kind of information system. It requires a correspondingly different evaluative framework, which this curriculum, and the ACRL AI Competencies it is built on, is designed to provide.

It is useful to hold two seemingly contradictory positions at once: AI tools are genuinely useful for library work right now, and they are also overhyped in ways that create real professional risks. Neither uncritical enthusiasm nor reflexive skepticism serves librarians well. Taken alone, each position prevents the calibrated professional judgment that good practice requires.

The claims made for AI in library contexts range from reasonable to extravagant. On the reasonable end, AI does help with drafting, editing, and explaining. On the extravagant end, AI will replace reference librarians, revolutionize cataloging overnight, or solve information literacy problems that decades of library instruction have not. For example, Clarivate's Pulse of the Library 2026 survey of 1,876 library workers found that about seven in ten libraries are exploring or implementing AI. Yet only about one in six has reached moderate or active implementation, and half of respondents are building their AI skills on their own rather than through institutional support. This gap between interest and practice is typical of early adoption periods. It argues for measured, evidence-informed practice rather than wholesale enthusiasm or wholesale resistance.

The ACRL AI Competencies framework identifies skepticism as a guiding mindset alongside curiosity, and that pairing is deliberate. We are meant to explore and question at the same time. This dual orientation is what librarians have always brought to information evaluation: the capacity to engage with new sources while keeping the critical apparatus that separates useful information from misleading information. AI does not require a new professional disposition. It requires applying an existing one to a new domain.

When introducing AI to skeptical colleagues, one useful framing treats it as an amplifier rather than a replacement. The professional value AI delivers depends directly on what the practitioner brings to the interaction: domain knowledge, institutional context, and an understanding of what a patron actually needs. A prompt written by a librarian who knows their collection, their student population, and their professional obligations produces a categorically different result than the same general prompt written by someone without that knowledge. This framing corrects an assumption shared by enthusiastic overclaiming and reflexive dismissal alike: that AI does something independent of the professional using it. It does not. The quality of AI output in professional library work is a function of the expertise the librarian brings to the conversation.

For developing this calibrated stance, the most useful frame is to treat AI as a capable but unreliable research assistant. You would use a capable assistant. You would also verify their work. You would not send the assistant's output to a patron without review. This frame guides practice more reliably than either enthusiasm or resistance alone, and it maps cleanly onto the professional standards library workers already hold.