Why the algorithm ignores your books

Recommendation algorithms appear to be objective, but in practice, they often unconsciously favor the most popular books in libraries. During her closing keynote at inct.spiratie 2026, researcher Savvina Daniil will reveal the hidden biases of these AI systems and show how the book industry can regain control of its own catalog.

The rise of artificial intelligence in our daily lives is now unstoppable. Seven years ago, when Savvina Daniil, a mathematician educated in Greece, came to Amsterdam to pursue a master’s degree in Data Science, AI was still far less ubiquitous. Even within her field, the emphasis at the time was primarily on optimizing the technical performance of systems.
“The focus was very much on accuracy,” Daniil explains in an interview with inct. How well can a system predict what someone will do or wants? According to her, that has since changed. Back then, scientists were already critically examining the societal implications of AI, but now that analysis is much more explicit: for whom does such a system actually work well, who is disadvantaged by it, and what choices underlie the outcomes?

Inclusivity

After graduating, Daniil worked for two years at TNO as part of a team focused on responsible AI. Their work centered on issues such as privacy, transparency, and the reliability of systems. She herself became particularly interested in inclusivity: do algorithms truly serve diverse groups of users, or do they primarily work well for the largest or most dominant group?

Culture, Ethics, and the Library

That interest eventually led Daniil to the Cultural AI Lab, a collaborative initiative in which academic research institutions and cultural organizations conduct joint research on artificial intelligence. Through that network, she began researching recommendation systems in the library sector, funded by the Royal Library and conducted at the Center for Mathematics and Computer Science (CWI).

According to Daniil, the reason was very specific. The KB recognized the potential of recommendation systems but also wanted to know how this technology could be used responsibly.
A key tenet from the KB’s ethical AI principles played an important role in this: anyone who wants to build inclusive systems must take bias in the underlying data into account. That bias must be mitigated or, at the very least, made visible. That became the starting point for the project “Responsible Recommenders in the Public Library Sector.”

An algorithm is not objective

One of the misconceptions Daniil is eager to dispel is the idea that algorithms are neutral calculators that automatically arrive at a single correct result.
In recommendation systems, however, uncertainty and probability play a major role, she explains. Systems must constantly make simplifications to generate a recommendation from enormous numbers of possible options.
During the conversation, Daniil uses a simple example. When an AI system has to recognize whether an image shows a dog or a cat, there is, in principle, one correct answer. When recommending books to readers, it’s different. There is never a single book that can objectively be identified as one a particular reader “must” read at a specific moment.

When a library system recommends a title and the user then borrows it, the algorithm may record that recommendation as a success. But, as Daniil emphasizes, that doesn’t tell us at all whether this was actually the best recommendation. There might have been thousands of other books that would have satisfied that user just as much.
Developers’ choices also influence the outcome. What data is used? Which factors carry the most weight? How is success defined? Change just one of those variables, and a completely different recommendation could result.

AI transforms ecommerce by enabling smart product recommendations, AI automated shopping decisions, and AI-powered ecommerce analytics to optimize customer experience and AI ecommerce success.
Photo: Shutterstock


Bias starts with the data

According to Daniil, bias doesn’t just arise within the algorithm. The data it uses is never completely neutral either.
“Data is constructed,” she emphasizes. After all, a book collection is the result of all sorts of prior choices. Which books were purchased? Which authors are published? Which titles have been promoted? Which books have been popularized by the media or social platforms?
For example, if a collection contains a relatively large number of American authors, a system that makes recommendations based on that collection will also recommend a relatively large number of American authors. The algorithm thus does not simply reflect “reality,” but also previous editorial, commercial, and cultural choices.

According to Daniil, it is precisely this that publishers, libraries, and booksellers need to realize: recommendation systems build on decisions that have already been made earlier in the chain.

Two ways to recommend books

During the interview, Daniil also explains how recommendation systems typically work. Broadly speaking, there are two approaches.
With content-based filtering, the system looks at a book’s characteristics: author, publication year, summary, subject, or, for example, difficulty level. It then compares that data with books the user has read before.
But even seemingly objective metadata contains human choices, says Daniil. A description such as “difficult” or “accessible,” for example, is not a physical quantity, but an interpretation.

The second method is collaborative filtering. In this approach, the system focuses primarily on the behavior of groups of users. If two readers have largely chosen the same books in the past, the algorithm can assume that they have similar preferences. If one reader then discovers a new book, it can also be recommended to the other.
According to Daniil, large commercial platforms nowadays typically use combinations of such techniques. But regardless of which model is used, people determine which signals are important and how much weight they carry.

The pitfall of popularity bias

Daniil’s own research focuses specifically on popularity bias. Under this bias, titles that are already popular automatically have an increasingly higher chance of being recommended again. This happens for a reason. From an algorithm’s perspective, well-known books are relatively safe recommendations. Users recognize the title and are therefore more likely to click on it or borrow it. On paper, this can even make a system appear exceptionally accurate. But according to Daniil, that is precisely where the problem lies.
A recommendation system that keeps showing the same popular titles has little added value. Libraries and booksellers, on the other hand, have extensive catalogs, many of whose titles remain largely hidden. According to her, the strength of recommendation systems should lie in connecting that so-called “long tail” to readers for whom precisely those lesser-known books might be of interest.

During her research, Daniil also found that popularity bias can exacerbate other forms of inequality. For example, popular books were found to come relatively often from authors in the Anglosphere, including the United States, Great Britain, Canada, and Australia.
An algorithm, therefore, does not need to take authors’ nationalities into account at all to develop a preference for certain groups. Simply giving extra weight to popularity can lead to such an outcome.

Bias isn't always wrong

At the same time, Daniil warns against automatically viewing bias as something negative. To determine whether a deviation is problematic, you must first identify the goal an organization aims to achieve with a recommendation system. During the interview, she cites the example of research conducted at Danish libraries. There, an algorithm was found to recommend books by Danish authors with striking frequency.
Technically speaking, that was a clear bias. But that doesn’t necessarily have to be a problem for the library. Promoting Danish literature was, in fact, one of the objectives. The situation changed, however, when Daniil observed that within that group, it was primarily well-known, popular Danish writers who stood out. As a result, debut authors and lesser-known writers received hardly any visibility. The system’s developers considered this undesirable.
So you always need some kind of reference point, Daniil explains. Only when it’s clear what you want to achieve with a system can you assess whether a particular bias is desirable or problematic.

Young Asian woman in CWI library, holding a book
Photo: Ivar Pel

A dashboard for the book industry

Daniil now wants to put that nuance to practical use. In her current research, she is working on a dashboard designed to give organizations insight into the behavior of recommendation systems. After all, not every librarian, publisher, or bookseller has the technical knowledge to fully analyze an algorithm. The dashboard is therefore designed to highlight any preferences or anomalies that occur within a system and their potential consequences.
The premise is explicitly not that a computer should then determine whether a system is “right” or “wrong.” Rather, Daniil wants professionals to be able to make that assessment themselves.
For example, a dashboard can show that a particular recommendation system highlights a relatively large number of popular titles, that local authors are becoming less visible, or that a certain group of users is being served less effectively. It is then up to the organization to determine whether that aligns with its own objectives.
According to Daniil, this shifts the discussion about AI from a purely technical issue to a strategic one: what values does your organization actually want to see reflected in its digital showcase?

Closing keynote at inct.spiratie 2026

The impact of recommendation systems extends far beyond the walls of the public library. For publishers and booksellers, there is a great opportunity—and responsibility—in understanding the algorithms that increasingly determine what content consumers see.
At inct.spiratie 2026, Savvina Daniil will deliver the closing keynote. She’ll take the audience behind the scenes of AI and, using concrete real-world examples, demonstrate how recommendations are generated, what biases can arise in the process, and—most importantly—what choices organizations can make for themselves.

This article was previously published on the INCT website, a platform for publishing experts.