The Closed-Book Effect

AI, the illusion of completion, and what librarians might build after chatbots

A student types a research question into an AI tool.

The answer appears quickly. It has a calm tone. It has sections. It uses the right vocabulary. It may even include citations. It does not look messy or unfinished, which is exactly the problem.

Before the student has searched a database, compared sources, checked a citation, opened an article, noticed a gap, or changed their mind, the work already feels partly done. Not because the research has been completed, but because the answer has the shape of completion.

That feeling deserves more attention.

Much of the conversation around generative AI and student research has focused on plagiarism, hallucinated citations, academic integrity, and prompt writing. Those are real issues. But I think there is another problem sitting slightly underneath them: AI can change a student’s sense of when research is finished.

I have started thinking of this as a closed-book effect.

The closed-book effect is what happens when a fluent AI-generated response makes inquiry feel settled before the student has opened the sources. It is not the same thing as a hallucination. A hallucination is an error in the answer. The closed-book effect is an error in the user’s relationship to the answer.

I think that distinction is critical to make.

An AI response can be partly accurate and still create the wrong feeling. It can give a useful overview and still make the student stop too early. It can cite real sources and still discourage the slower work of reading them carefully. The danger is not only that AI gives students bad information. The danger is that it gives them the feeling of being done.

This is where librarianship has something to say.

A good reference interaction rarely begins and ends with the first question asked. Someone comes to the desk and says they need sources on climate change, or prison reform, or artificial intelligence, or Toni Morrison, or urban planning. A librarian does not simply hand over the first plausible answer. The work is more delicate than that.

What kind of assignment is this? What level of source is needed? Is the student looking for background, argument, data, criticism, history, or a position to respond to? Are they trying to understand a topic, defend a claim, compare perspectives, or find a gap in the literature? Have they searched yet? Where? What terms did they use? What did they find? What confused them?

This is not just customer service. It is intellectual orientation.

The reference interview is one of the most underrated models of human help. It assumes that the first version of a question is often incomplete. It treats confusion not as a defect, but as part of the research process. It gives the user a way to move from a vague need to a more exact path.

That is very different from the usual AI interface.

The dominant AI interface is still the answer box. You ask. It replies. You ask again. It replies again. There is nothing inherently wrong with this. It can be useful, especially when the user already knows what they are doing. But for students who are still learning how research works, the answer box can be misleading. It can make research look like a sort of transaction instead of an investigation.

Research is not only the production of an answer. It is a change in the researcher.

A student begins with one idea of the topic and, if the process is working, leaves with a more complicated one. They learn new terms. They find authors they did not know existed. They discover that scholars disagree. They realize that the first question was too broad, or too narrow, or built on a shaky assumption. They learn which sources count in a discipline and which sources only sound authoritative. They learn how knowledge is made.

The ACRL Framework for Information Literacy already gives us language for much of this. Authority is constructed and contextual. Research is inquiry. Scholarship is conversation. Searching is strategic exploration. Generative AI puts pressure on each of these ideas.

It can flatten authority by presenting claims in a single confident voice. It can make inquiry look complete. It can simulate scholarly conversation without requiring the student to encounter actual scholars. It can short-circuit searching by giving the user a neat paragraph where a messy trail of sources should be.

This does not mean librarians should reject AI tools. That would be too easy, and probably not useful. Students are already using them. Faculty are using them. Library vendors are building them into discovery systems. The real question is not whether AI belongs near research. It is what kind of research habits survive contact with it.

I think the answer depends less on prompt quality than judgment quality.

Prompt quality asks, “How do I get the tool to give me a better answer?”

Judgment quality asks, “How do I know what this answer has earned?”

That second question feels more durable. The tools will change. Interfaces will change. Some citation problems will improve. Models will become better at retrieval. But students will still need to decide what to trust, what to check, what to read, what to doubt, and what to do next.

A student evaluating an AI-generated research response has to make several judgments at once.

Is the claim true? What evidence supports it? Do the citations exist? Do the sources say what the AI claims they say? Are the sources appropriate for the field and the assignment? What perspectives are missing? What is being overstated? What would a skeptical reader ask? What should I search next?

That last question is the one I keep returning to.

What should I search next?

The best student researcher is not the one who receives the most polished AI response. The best student researcher is the one who can look at that response and say, “This is useful, but it has not yet earned my trust. Here is what I need to check. Here is what I still need to read. Here is where the uncertainty is.”

That kind of student is not anti-AI. They are not naïve or gullible either. They have learned how to keep inquiry open.

I think librarians are especially well positioned to teach this because the library has never only been about finding things. It has also been about helping people understand what kind of thing they have found.

A scholarly article is not just “better” than a blog post in every situation. A primary source is not useful simply because it is old. A citation is not valid simply because it appears in a bibliography. A source can be peer-reviewed and still limited. A database search can be technically successful and intellectually poor. Authority shifts by context. Search terms carry assumptions. Absence in search results does not always mean absence in the world.

These are ordinary librarian insights. In the age of AI, they become more urgent.

This concept of a closed-book effect also makes me wonder whether the standard chatbot interface is the wrong metaphor for serious inquiry.

A chat window encourages an exchange: question, response, follow-up, response. It can be efficient, but it does not naturally show the user the structure around a topic. It does not feel like browsing shelves, comparing editions, following citations, moving from background reading into criticism, or discovering an adjacent field by accident. It collapses too much into the answer.

A library does something else.

A library gives knowledge a shape. You can stand in front of a shelf and see that a book has neighbors. You can notice that a subject has depth. You can wander from one call number range into another. You can enter looking for one thing and leave with a better question.

That experience is not nostalgic. It is a design principle.

What would an AI system look like if it were built from that principle instead of the answer box?

I imagine something closer to a virtual reference library.

The user would not begin by asking for a final answer. They might begin with a title, a topic, a research question, or even a confession: “I am reading The Souls of Black Folk and I do not know what I should understand first.” Or: “I am trying to get oriented in urban planning.” Or: “I need to write about AI in higher education, but everything I find feels either too technical or too shallow.”

Instead of returning a single polished response, the system would build a guided space around the inquiry.

One area might provide essential context. Another might show historical background. Another might gather major debates. Another might offer companion essays, lectures, interviews, or videos. Another might show related books and unexpected neighboring subjects. Another might offer discussion questions. Another might end with only a few next steps, because too many recommendations can become another form of noise.

The point would not be to make research cute. I am not interested in badges, streaks, or fake academic gamification. The point would be to create an interface that makes inquiry feel open, structured, and alive.

In library terms, the central object might be a Pathfinder.

Pathfinders are not final answers. They are guides for beginning. They help users locate important sources, terms, debates, and directions. They are humble by design. A good Pathfinder does not pretend to exhaust a subject. It says: here is enough structure to begin without being lost.

That is exactly the posture AI research tools need more of.

A Pathfinder for a book, topic, or question could include the basics: essential context, key themes, historical background, major debates, companion essays, lectures and videos, related books, discussion questions, adjacent shelves, and three recommended next steps. But the deeper purpose would be to resist false completion. Every section would remind the user that understanding is something built through movement, not downloaded in one response.

There is also a privacy argument here.

Some visions of AI assistance lean toward ambient monitoring: tools that watch, listen, infer, remember, and personalize everything. That may be useful in some settings, but it feels wrong for a library-minded system. Intellectual privacy is not a decorative value in librarianship. It is part of the ethics of the field.

A virtual reference library should not need to follow a user around the web or quietly absorb their life. The user should choose what to bring into the system. A question. A title. A draft. A topic. A source list. A syllabus. A problem they are trying to understand. The system should respond to that chosen inquiry, not build a hidden profile from the user’s behavior.

The library metaphor helps here too. You can walk into a library and ask for help without surrendering your entire inner life. The help is situated, respectful, and bounded. An AI system inspired by libraries should inherit that restraint.

This is part of what I am now beginning to explore through a broader idea: a virtual reference library for helping people find their way through what they are reading, studying, researching, or trying to understand. Not an answer machine. Not a surveillance tutor. Not a productivity dashboard with nicer fonts. A place for orientation.

That word, orientation, keeps doing a lot of work for me.

Students do not only need information. Readers do not only need summaries. Researchers do not only need sources. Very often, people need to know where they are, what kind of thing they are looking at, what is missing, what matters, what is disputed, and what they should do next.

This is the older work of librarianship, and it may also be one of the more useful design models for AI.

The future of AI in research should not be measured only by how quickly a system can answer. Speed is not the same as understanding. Fluency is not the same as authority. A citation is not the same as evidence. A finished-looking paragraph is not the same as completed inquiry.

The closed-book effect names a problem, but it also points toward a design opportunity.

If AI can make students stop too early, then library-minded AI should help them continue well. It should slow the moment of closure just enough for judgment to enter. It should ask where the evidence is. It should show the nearby shelves. It should make disagreement visible. It should treat uncertainty as useful information. It should help the user leave with a better next move, not just a smoother answer.

Maybe the next important AI interface is not a smarter box.

Maybe it is a room.

A room with a reference desk, a few open books, a shelf of neighboring questions, and a subtle reminder that the first answer is rarely the end of the search.

References:

Association of College and Research Libraries. (2016). Framework for information literacy for higher education. American Library Association. https://www.ala.org/acrl/standards/ilframework

Association of College and Research Libraries. (2025, October). AI competencies for academic library workers. American Library Association. https://www.ala.org/acrl/standards/ai

Breeding, M. (2025, May 1). 2025 library systems report: Companies see platform upgrades, new leadership, and AI enhancements. American Libraries Magazine. https://americanlibrariesmagazine.org/2025/05/01/2025-library-systems-report/

International Federation of Library Associations and Institutions. (2025, April 6). IFLA statement on copyright and artificial intelligence (AI). IFLA Repository. https://repository.ifla.org/handle/20.500.14598/3927