Gains:
- Ability to distinguish between content-based and collaborative recommendation approaches and prevent fake sources by verifying each recommendation in the real collection
- Ability to break the filter bubble with conscious diversity and present suggestions with transparent justifications
- Protecting the privacy of user reading and search history and processing this data securely
A reader might ask "I finished this book, what should I read next?" ', a good librarian recommends resources appropriate to his or her interest, level and purpose. In the digital world, recommendation systems do this at scale: they recommend new resources by looking at a user's past interest or the characteristics of a resource. In this unit, you will learn how artificial intelligence-based recommendation systems work, what they bring to the library, and how to manage risks such as privacy, bias and "filter bubble".
Let's define the concepts. Content-based recommendation looks at a source's characteristics (topic, author, genre) and recommends similar sources: "If you liked this history book, you might also like this one, which covers the same period." Collaborative filtering works on the principle of "other users who liked resources like you also liked these". A filter bubble is when a recommendation system locks the user into constantly similar content and keeps them away from new, different perspectives. A good library recommendation both captures interest and breaks the bubble.
Step by step: a responsible advice approach
1. Clarify the purpose. What does the proposal aim for? Do you want the user to continue reading, broaden their horizons, or find resources for a specific research question? The purpose determines the type of recommendation.
2. Verify source and authenticity. AI can make up non-existent books when making suggestions. Verify that each suggested resource actually exists in the collection or in a reliable catalogue.
3. Maintain diversity. Recommending only the "most similar" one creates a filter bubble. Consciously include different perspectives, different authors, and different levels. The task of the library is to open horizons, not to confine them to the ecology.
4. Protect privacy. Personalized recommendation uses the user's reading history and is extremely accurate. Do not provide this data to general AI tools; Process only within the institution's secure systems, with the user's consent.
5. Be transparent. Explain to the user why the recommendation was made: "This resource was recommended because it is related to the topic you are interested in." Transparency strengthens trust and information literacy.
Tip: Consciously add a “slightly different” resource to a list of suggestions and state: “These are close to your area of interest; I added this one as a refresher.” This both breaks the filter bubble and demonstrates the discovery value of the library.
Ethical dimension of the proposal
Recommendation systems seem innocent but create powerful effects. Which resources are recommended to a user shapes his or her information world. If the system only highlights sources that are popular or only from a particular point of view, minority voices, differing views, and less well-known valuable sources will be invisible. AI can reinforce this trend by inheriting the popularity and biases in the training data.
This is where the value of librarianship comes in: the library is not a commercial recommendation engine; Its mission is to ensure equal and diverse access to information. It is the librarian's responsibility to align the recommendation system with this value.
Attention: User's reading and search history is one of the most sensitive privacy data. What a person reads can reveal information about his or her health, political views, beliefs, or personal crisis. Never transfer this data with its identity to external systems; Anonymization and consent are essential.
three mini cases
Case 1 — Content-based recommendation worked. A school library used a content-based system to recommend resources to students that were similar to the book they were reading. One student finished a science fiction novel; The system recommended three books covering the same theme but from a different author. The student borrowed two; The reading habit continued. The librarian had previously confirmed that all of the suggestions actually existed in the collection.
Case 2 — Fake suggestion caught. A librarian asked an AI to recommend 8 books to a user. He checked the list in the catalogue: 3 out of 8 books were missing at all; The AI had made up plausible headlines. The verification step prevented non-existent books from being recommended to the user.
Case 3 — Filter balloon ruptured. A researcher was constantly working with sources with the same point of view, and the system always suggested similar ones. The librarian deliberately included two sources that presented a contrasting approach to the topic and made this clear. It was only thanks to these opposing sources that the researcher realized the incomplete aspect of his study.
Cold start and new user issue
A well-known challenge of recommender systems is the cold start problem: the system does not know what to recommend to a new user or a resource that has just been added to the collection for which there is no historical data. In this case, many systems turn to the most popular resources as the safe option; This means that little-known valuable resources are not recommended at all and the popularity feeds on itself. In librarianship terms, this conflicts with the value of equal access. The solution is to ask the new user a short, clear preference question (which topics, which level, which purpose) before making a recommendation and highlighting newly added resources based on their content characteristics. At this point, by analyzing the subject and features of a resource, AI can link even a book that has never been borrowed to the relevant user; as long as the suggestion comes from the real collection and is verified. Thus, the system makes visible not only the "very loved" but also the "relevant and valuable".
Tip: When setting up a recommendation for a new user, base it on the user's current needs rather than the system's historical data. One good question (“is this for research, for pleasure, to what degree?”) often produces a more accurate recommendation than months of behavioral data.
Four copyable templates
1) Content-based similar resource recommendation:
Below is the subject/summary of a resource that a user liked. I have also given a list of existing resources in our collection. Just from this list, suggest 5 resources similar to the subject and write a one-sentence justification for each. Suggesting sources not listed. Favorite source: [here] / Collection list: [here]
2) Suggestion that adds variety:
To the list of suggestions below, add 2 resources that will broaden the user's horizons and offer a different perspective or subtopic (only from the collection list I provided). Explain why they are different and valuable. List: [here] / Collection: [here]
3) Explanation of the reason for the recommendation:
For each recommendation below, write a short, transparent justification that can be shown to the user: "This resource was recommended because..." No exaggeration, no implication of personal data. Suggestions: [here]
4) Privacy pre-check:
Does the following suggestion entry contain user-identifying or sensitive information (insinuating health, beliefs, political views)? If so, mark it and suggest how to anonymize it. Input: [here]
Weak prompt / Strong prompt
Weak prompt:
Recommend 10 books for this user to read.
The AI generates suggestions from its global memory without knowing the collection; He can make up books that do not exist and does not respect diversity and privacy.
Powerful prompt:
Your role: reader advisory assistant. Below is the topic the user liked and the list of real sources in our collection. (1) Suggest 5 sources related to the topic from the list. (2) Also add 2 sources with different perspectives that will open your eyes. (3) Write a transparent justification for each suggestion. Do not go off the list, fabricate sources, or imply personal data. Subject: [here] / Collection: [here]
Powerful prompt; it limits it to the actual collection, demands diversity, and adds justification and a privacy rule.
Recommendation type comparison chart
Suggestion type
strong point
risk
Librarian's duty
content based
Captures attention well
filter bubble
Add variety
collaborative
surprise discovery
Privacy, popularity bias
Anonymity, balance
AI production list
fast
made up source
verify reality
Common mistakes
- Offering suggestions without verification. AI can make up non-existent sources.
- Just recommending "most similar". The filter bubble narrows the user's horizons.
- Exporting reading history to external tool. It is the most sensitive data of privacy.
- Presenting the suggestion without justification. Transparency supports trust and literacy.
- Ignoring popularity bias. Valuable but little-known resources are lost.
In summary
Recommendation systems are powerful tools that facilitate discovery by capturing user interest; Content-based and collaborative approaches carry different strengths and risks. AI is fast at generating suggestions, but can invent non-existent sources, filtering bubbles and reinforcing bias. The job of the librarian; To verify every suggestion in the real collection, to break the bubble with conscious diversity, to protect user privacy and to maintain the value of equal and diverse access of the library by presenting suggestions with transparent reasons.
Application task
Prepare a short list of factual collections (10-15 sources). Generate a recommendation for a user with the "content-based similar resource recommendation" template and verify that they all actually exist in the list. Then, add sources with different perspectives to the list with the "Suggestion that adds diversity" template. Finally, create a transparent explanation for each suggestion with the "Suggestion rationale explanation" template and plan how you will present it to the user.
checklist
- [ ] I verified that each suggestion actually exists in the collection.
- [ ] I added variety to break the filter bubble.
- [ ] I kept the user reading/search history safe, I did not give it to an external tool.
- [ ] I have prepared a transparent justification for each suggestion.
- [ ] To avoid popularity bias, I kept an eye on lesser-known and valuable sources.