Gains:
- Being able to distinguish where artificial intelligence saves time in the librarianship workflow (metadata, summary, search, suggestion) and where accuracy, source reliability and privacy decisions are left to humans, depending on the level of risk.
- Ability to apply a discipline that verifies each output by linking it to the source, confirming it in an independent source, and passing it through a bias filter.
- Understanding why hallucination, fabricated sources and user privacy are risks that must be taken into account in this profession from the very beginning.
When a cataloging specialist sits down at his desk in the morning, he may have dozens of newly arrived books, hundreds of documents from a donation collection, and an archive box waiting to be digitized. A reference librarian (the expert who helps users find information) answers questions throughout the day like “where is the reliable source on this topic?” An information manager is responsible for keeping an organization's thousands of documents discoverable and reliable. In this unit, we will clarify where exactly artificial intelligence (AI in short; large language models and similar tools) saves time in these tasks, and where judgment, choice and responsibility remain with the human.
Let's start with a definition: artificial intelligence, by which we mean language-generating and pattern-recognizing software tools trained on large amounts of text. These tools summarize, translate, suggest classification, generate draft text, and flag patterns in large chunks of data. But it does not know whether a source really exists, whether an information is correct, or whether a user's privacy is protected. That's why in librarianship and information management, AI is an assistant, not a decision maker.
Where does artificial intelligence work and where does it not work?
Let's divide the librarianship workflow into three layers. The first layer is mechanical and repetitive tasks: extracting draft metadata from the byline of a book, summarizing a long report, translating text from one language to another, cleaning a table of data. AI is powerful here and saves big time.
The second layer is quasi-judicial work: what subject heading to give a document, whether to include a resource in a collection, what resource to recommend to a user. AI produces suggestions here, but the expert who knows the institution's policy and the real needs of the user has the last word.
The third layer is pure judgment and accountability: confirming that information is accurate, deciding on a copyright, whether to disclose sensitive personal data, confirming the reliability of a source. At this layer, the AI's output is just a starting point; The responsibility lies entirely with humans.
Tip: Before giving a task to the AI, ask yourself: “What happens if this output is wrong?” If the answer is "I'll waste some time", feel free to use AI. If the answer is "a false source is spread, a user is misled, or personal data is leaked", be sure to verify the output.
Hallucination: the librarian's biggest danger
Hallucination is when AI produces non-existent information, source or quotation as if it were real, and in an extremely convincing language. This is particularly dangerous for librarianship and information management; because the essence of our profession is the authenticity and reliability of the source.
When you tell the AI to “suggest five academic papers on this topic,” it can generate titles, authors, and journal names that seem real but don't exist at all. These fabricated sources are so convincing that a patron or even a librarian can add them to a bibliography without verifying them. In the humanities and social sciences, the danger is even greater: when "that document from 1897" is fabricated for a historian or "that court decision" for a jurist, the result is serious misinformation.
Attention: No source citation, quote or reference produced by AI should be transferred anywhere without verification in a real catalogue, database or the source itself. Just because it "looks convincing" doesn't mean it's true.
Verification discipline: three steps
The basic discipline that the librarian must apply when working with AI is this:
1. Connect to source. Every fact, figure, date and imprint must be based on an original source. Ask the AI “tell me from what source you got this information” and check that source independently.
2. Independently confirm. Verify critical information in at least one reliable source outside the AI (authentic catalogue, official database, primary document).
3. Pass it through the bias filter. AI carries the biases of the data it is trained on. A suggested topic, a list of suggestions, or a summary may highlight certain points of view and obscure others. Ask who is left out, what perspective is missing.
Privacy and user privacy
One of the most fundamental ethical principles of the librarianship profession is user privacy: what a person reads, what he searches for, and what subject he is interested in are confidential. Entering data into AI tools, especially public tools that operate over the Internet, that includes users' identities, search history, or borrowed resources may violate this principle; because this data may go to third party systems.
Likewise, in an organization, unclassified but sensitive documents (personnel information, contracts, internal correspondence) should not be uploaded to a general AI tool. Tools running on the institution's own secure, closed system are preferred for this type of data.
Tip: Before entering any data into the AI, ask "Is this information public?" ask. If the answer is no, or you're not sure, put that data in a de-personalized, anonymous format or use only secure tools approved by the institution.
three mini cases
Case 1 — Draft metadata for 400 documents. A university library had to catalog an archive of 400 documents donated by a professor. The expert extracted a draft title, date and topic suggestion from the first page of each document with AI: average time per document decreased from 12 minutes to 4 minutes. But each draft imprint was compared with the original document and approved by the expert; 17 incorrect dates made up by the AI were caught in this step.
Case 2 — Fake source caught. An advisory librarian received 6 article suggestions from YZ while searching for resources on "climate migration" for an undergraduate student. The librarian searched them all in the institution's database: 2 out of 6 articles did not exist at all. The verification step prevented the student from being given a fabricated source.
Case 3 — Confidentiality breach prevented. A library considered using a general AI tool to analyze “which user is interested in which topic” from checkout records. The information manager realized that it would be a violation of privacy for this data to go out with the user ID; the data was first anonymized and processed solely as aggregated, de-identified statistics.
Four copyable templates
1) Start defining roles and boundaries:
Your role: experienced assistant cataloging and information management specialist. Your task: to produce a draft from the information in the source I gave you. Rules: (1) Only use the information in the text I gave you, do not add your own information. (2) Mark "[must be verified]" where you are not sure. (3) Do not make up any dates, names or imprints. If you understand, start. Source: [here]
2) Validation check question:
List each factual claim (date, name, number, source citation) in the text below. For each: where in the text did you get this information? If it is not clearly in the text, write "not in the text". Text: [here]
3) Bias and missing perspective checking:
Which perspectives does the following summary/recommendation highlight and which does it leave out? What groups, regions or views might be missing? Just rely on the text, don't speculate. Text: [here]
4) Privacy pre-check:
Does the text below contain personal data (name, ID number, contact, reading/search history) or sensitive corporate information? If so, list it and suggest how it can be anonymized. Text: [here]
Weak prompt / Strong prompt
Weak prompt:
Find me some resources on this subject.
This prompt invites the AI to generate fake resources. There are no roles, boundaries, validation rules and resources. AI can produce believable but non-existent tags.
Powerful prompt:
Your role: reference librarian assistant. Choose the ones related to the subject of "climate migration" from the catalog entries I have given you below and explain why they are relevant. Do not add any sources that are not on the list, do not fabricate sources. When you're not sure, say "there are no suitable sources listed." Catalog entries: [here]
The difference is clear: strong prompt forces the AI to work from the given real source, not from its own memory, and prevents fabrication.
Layer and responsibility table
business layer
sample task
Role of AI
man's decision
mechanical
Draft metadata, summary, translation
Produces drafts quickly
verification, approval
quasi-judicial
Topic title, source suggestion
offers suggestions
Selection by policy
pure judgment
Authentication, copyright decision
Preliminary information only
Full responsibility lies with the person
Common mistakes
- Substituting the AI output for the source. AI is an assistant, not a resource; Every output is validated.
- Accepting the fabricated source as convincing. Just because it looks good doesn't mean it's right.
- Entering personal/sensitive data into the public tool. User privacy and corporate confidentiality are respected from the very beginning.
- Ignoring prejudice. AI carries the perspectives of the data on which it is trained; Ask what's missing.
- Leaving the judgment to AI. Decisions of accuracy, copyright, and privacy belong to humans.
In summary
artificial intelligence; It is a powerful assistant for summarizing, drafting, translating and pattern marking in librarianship and information management. It saves a lot of time in mechanical work; gives advice on quasi-judicial matters; But verification of accuracy, reliability of the source, user privacy, and ethical decisions belong to humans. Hallucination is the greatest danger to this profession: no source credits or quotes are used without verification. Run AI with clear role, boundary, and validation rules; Confirm each critical output with an independent source.
Application task
Choose a real topic from your field. Ask the AI for 5 resource recommendations on that topic, then search each one in an actual catalog or database and note how many actually exist. Then, with the “Startup defining roles and boundaries” template, tell the AI to work only from a text you provide and compare the output of the two approaches. Write which approach prevents fitting.
checklist
- [ ] I determined which layer (mechanical, semi-judicial, pure judgment) the task is at.
- [ ] "What happens if this output is incorrect?" I asked the question.
- [ ] I have verified each credit and citation against an independent source.
- [ ] I have checked that there is no personal/sensitive information in the data entering the AI.
- [ ] I kept the final decision and responsibility at my own discretion.