What Happens to Serendipity When Algorithms Choose What We Read?

Geoffrey Chen

How a Person Builds a World · Season One: “Reading, Memory and Judgment” · Article Ten

One reader visits a bookshop looking for history and, after turning into the wrong aisle, picks up a collection of poetry. Another waits at a library desk and notices an unfamiliar author on the returns trolley. A third opens “Recommended for you” on a reading platform and finds a book closely related to previous purchases. All three encounters might be called serendipity, but their accidents are not the same.

The first two emerge from spatial proximity, another reader’s return and the edge of a classification. The third is produced by a system that has calculated relevance from earlier behaviour. An algorithm can also reveal a genuinely unexpected and valuable book, but it usually begins by turning “what you may like” into an objective that can be optimised.

Recommendation algorithms have not abolished chance. They have changed where chance comes from, who arranges it and which kinds of surprise are most likely to appear.

Discovery was never purely random

Nostalgia for a pre-algorithmic reading world ignores older forms of selection. Bookshops chose which titles to stock and where to display them. Libraries used classifications and subject headings. Newspaper editors selected books for review. Teachers, friends and prizes established reading paths. The chance of wandering into a book was already organised by architecture, market and cultural authority.

Those older systems were not inherently fair. Limited entrances, publishing concentration and weak visibility for minority languages are longstanding problems. Recommender systems can lower the cost of searching very large catalogues. They may deliver a small or specialised work to a reader who would otherwise never encounter it.

The issue cannot be reduced to authentic human recommendation versus artificial algorithmic recommendation. Both organise visibility. What distinguishes contemporary personalisation is its capacity to use individual behaviour as input, update quickly and feed each click back into the next recommendation.

Predicting preference makes the past an entrance to the future

Most personalisation uses some form of similarity: this item resembles what you previously consumed; people with related behaviour chose it; or it is likely to be selected in the present context. Real systems are much more complicated, but the common direction is to infer a next step from existing signals.

This is useful, but it establishes a temporal structure. The past is not merely recorded. It becomes a principal gate into the future. Read one detective novel and more detective fiction appears. Click several essays expressing a position and similar material receives more exposure. A further click gives the system stronger evidence. Preference and recommendation form feedback.

This feedback does not inevitably imprison every user inside a fixed “filter bubble”. Findings vary by platform, definition, metric and period. A systematic review of filter bubbles in recommender systems emphasises the inconsistency of concepts and evidence. An earlier longitudinal study used the semantics of user-generated tags to examine recommendation and content diversity over time, illustrating why the issue must be studied as a process rather than a single list. See “Exploring the Filter Bubble”.

The more defensible concern is not that algorithms necessarily close a person’s world. It is that the user cannot see the counterfactual: what might have appeared on this page if the system had not predicted from the past?

Surprise is not improved simply by becoming random

In recommender research, serendipity usually includes more than surprise. An item should also have relevance or value. A completely random book list is unexpected but may be useless. A genuine encounter is with something not deliberately sought that later proves important.

A useful discovery mechanism therefore works between familiarity and strangeness. Too much familiarity repeats preference. Too much strangeness provides no bridge. A book about urban planning might reach a reader interested in public space while leading into an unsearched history of architecture. There is continuity and deviation at once.

Some book recommender research has explicitly tried to increase serendipity. One study used linked data to find cross-domain book relations that were not obvious but remained explainable. See “Book Recommender System Using Linked Data for Improving Serendipity”. The work demonstrates that algorithms need not only provide more of the same. Diversity, novelty and serendipity can also become design objectives.

Quantifying those objectives creates another difficulty. How can a system know that a book changed a reader’s thought years later? Clicks, dwell time and purchases are readily recorded; long intellectual formation is not. A platform therefore tends to optimise measurable immediate behaviour rather than an unmeasurable process of becoming.

Reading interest should not be defined only by behaviour

What a person clicks is not identical to the reader they hope to become. Repeatedly selecting light material while tired is an immediate behaviour. Committing a year to understand a difficult thinker is a longer aspiration. Behavioural data can observe the first more easily than the second.

When a recommender infers preference entirely from previous action, it describes the user as the self already performed. A knowledge life also consists of aspiration, duty and deliberate self-change. A reader may wish to encounter another political position, another language or an ignored history even when initial click-through is low.

Readers therefore need to tell a system more than “like” and “dislike”. They should be able to say: I want to expand in this direction; do not let this one behaviour continue to shape my profile; show material further from the present pattern. Research on user control has found that more transparent interfaces can help people recognise the character of their information environment. See the study “How Does Empowering Users with Greater System Control Affect News Filter Bubbles?”.

Meaningful control is not a reset button buried in settings. It includes seeing reasons for recommendation, editing history, changing the degree of exploration and choosing an entrance without personalisation.

Some discovery spaces should refuse complete personalisation

A public library display, an independent bookseller’s choice, a course list and a friend’s recommendation are not automatically better than an algorithm. They bring different sources of judgement. A healthy reading environment should not be monopolised by one recommendation logic.

Several kinds of discovery space can coexist:

  • Continuity: deeper routes through a current interest.
  • Adjacency: movement from one subject into a neighbouring field.
  • Heterogeneity: deliberate exposure to other languages, periods and positions.
  • Public selection: the same material visible to everybody, independent of personal history.
  • Randomness: an entrance that allows genuinely purposeless selection.

No one space should determine reading. Their coexistence allows movement among efficiency, exploration and shared public experience.

The public space is especially important. If every home page is entirely different, we lose not only individual chance but common chance: many people meeting the same unfamiliar work on a shared shelf and thereby gaining an object they can discuss together.

A reader must complete the accident

However a book appears, display alone does not complete discovery. An unexpected book becomes a genuine encounter only when the reader pauses, tolerates initial unfamiliarity and gives it time. An algorithm can introduce deviation; it cannot guarantee the reader will accept it.

Attention is therefore a condition of serendipity. An interface built for continuous scrolling and immediate judgement may present diverse material while allowing every unfamiliar item to disappear in seconds. A finite list, a short explanation or a physical volume that can be opened gives the unfamiliar object a longer existence.

From a Sustenesis perspective, a reading world forms through preference, platform objectives, classification, interface, friends, institutions and readerly commitment. An algorithm is not an external invader but an increasingly powerful set of relations within that structure. The question is not how to remove it. It is how to prevent measurable behaviour from becoming a person’s entire future.

The deepest way recommendation changes accidental reading is not by making everything predictable. It quietly decides which forms of unpredictability retain a chance to appear. A trustworthy reading system should understand the past while preserving the right to depart from it. It should improve relevance and maintain spaces that have not been optimised away.

A knowledge life cannot ask only, “What am I most likely to enjoy next?” It also needs a question no behavioural history can answer for us: which unfamiliar thing am I willing to let change me?

Primary sources and further reading

Continue reading: Explore the How a Person Builds a World series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.