Recommendation algorithms might be making your entertainment boring, new research suggests
Eric W. Dolan · article · 2026-06-02 source ↗ Reader 🖥️
Algorithmic Curation & Taste NarrowingGoodhart's LawProxy Validity (Construct Validity of Measurement Proxies)
Because algorithms that optimize for short-term engagement mistake familiarity-driven engagement for quality and evaluate over a horizon far shorter than the decade-plus timescale on which aesthetic taste actually evolves, they systematically under-explore and trap users in an eventually-boring set of favorites — and a mathematical model shows that deliberately injecting moderate noise (forcing exploration) into such a system improves users' long-term satisfaction.
theoretical · expository — Science journalism reporting a theoretical economics paper (Samsun Knight, Journal of Cultural Economics). The underlying study is a pure mathematical model verified by Monte Carlo simulation, not an empirical study of real users — no human participants, no real-platform data. The article is faithful about this limitation in its final third, but the headline and framing lean harder toward a general claim about real recommendation algorithms than the model itself licenses. Practical implications (how platforms could redesign curation) are gestured at in the last two paragraphs but not developed.
Structure
| Part | Role |
|---|---|
| Knight's personal and professional anecdotes (Spotify recommendations, publishing-industry trend-chasing) plus the Bourdieu reference that sparked the idea | motivation / context |
| Explanation of 'consumption capital' — the inverted-U curve of appreciation from moderate vs. excessive exposure — and the hip-hop counterfactual | conceptual groundwork |
| Description of the two-part mathematical model (appreciation dynamics + algorithmic curator) and the Monte Carlo verification method | methodology |
| Findings: exploration rate converges to zero, the 'straddling' failure mode, and the self-fulfilling prophecy of monotony | evidence |
| Findings: moderate noise improves long-term satisfaction, especially with 3+ items | evidence / core counterintuitive claim |
| Caveats (theoretical-only, no real-user data, real-world testing impractical) and future-research suggestions | limitation / qualification |
| Closing suggestion that platforms track familiarity as a changing, multi-year state | practical implication (underdeveloped) |
Key terms
- consumption capital Algorithmic Curation & Taste Narrowing
The appreciation a person builds for an artistic style through repeated exposure — but only up to a point: moderate exposure increases liking, while excessive exposure causes boredom or satiation. An inverted-U, not a monotonic curve. - straddling
The failure pattern where an algorithm bounces between two poor choices: overexposing a genuinely good item until the user is sick of it, and just barely exposing a low-quality item enough to confirm it's not good — never discovering that resting the good item and revisiting later would restore enjoyment. - exploration rate
How often the simulated algorithm shows a user something unfamiliar/untested rather than a known favorite; the model tracks this as the key dependent variable and shows it converges to zero under short-sighted engagement optimization. - flawed algorithm (vs. the taste-aware algorithm)
Two simulated curator types Knight compares: one that mistakes high engagement for high underlying quality without realizing its own past recommendations caused the familiarity; another that correctly models changing taste but still only optimizes over a short horizon. Both converge to under-exploration, for different reasons.
Propositions
- explicitHuman appreciation for an artistic style follows an inverted-U curve: moderate repeated exposure increases liking ('consumption capital'), but excessive exposure causes boredom or satiation.conf: high
- explicitAn algorithm that mistakes high engagement for high underlying quality, without recognizing that its own past recommendations caused the familiarity driving that engagement, will treat a user's declining interest in unfamiliar content as proof the content is inherently bad.conf: high
- explicitUnder this dynamic, an engagement-optimizing algorithm's exploration rate mathematically converges to zero, closing it off entirely to new content and producing a self-fulfilling prophecy of user boredom.conf: highextendsThe 8 software engineering metrics AI broke · p7 — Generalizes the same Goodhart's-Law dynamic — a short-term, easy-to-measure proxy (engagement) displacing the true, hard-to-measure long-horizon target (durable satisfaction/quality) — from AI-era engineering metrics to consumer recommendation-algorithm design, a different domain entirely.
- explicitThis produces 'straddling': the algorithm overexposes a genuinely good item until the user tires of it, and underexposes a mediocre item just enough to keep confirming it's mediocre — never discovering that resting the good item would restore enjoyment.conf: high
- explicitEven an algorithm that correctly models taste as changing over time still under-explores in practice, because its evaluation window (weeks to months) is far shorter than the multi-year timescale over which aesthetic tastes actually evolve.conf: high
- explicitDeliberately introducing moderate prediction error (noise) into a recommendation algorithm increases users' long-term satisfaction, because the resulting accidental exposure to unfamiliar content lets them build appreciation they would otherwise never accumulate.conf: high
- explicitA recommendation algorithm that is extremely good at predicting what a user wants right now can, over years, narrow rather than expand the total range of things that user is ever capable of enjoying.conf: high
- implicitThe emergence of genuinely new genres (the article's example: hip-hop's initial reception as abrasive to rock-and-roll listeners) may historically have depended on curators tolerating low short-term engagement long enough for audiences to learn to appreciate the new style — something a pure engagement-optimizing algorithm would not do.conf: medium
Arguments
Highly accurate, engagement-optimized recommendation algorithms degrade users' long-term entertainment satisfaction, even when the platform sincerely wants to please them. strength: medium
Explicit premises:
- Appreciation for a style requires moderate repeated exposure to build, but too much exposure produces boredom (the inverted-U 'consumption capital' curve).
- Algorithms optimizing for short-term engagement read low initial engagement with unfamiliar content as evidence the content is low-quality, not as an artifact of unfamiliarity.
- This causes exploration rate to converge to zero, trapping users in 'straddling' between overexposed favorites and permanently-dismissed unfamiliar content.
- Even taste-aware algorithms fail if their evaluation horizon (weeks/months) is shorter than the multi-year horizon over which tastes actually evolve.
Implicit premises:
- Short-term user engagement (clicks, plays, immediate ratings) is not a valid proxy for a user's long-term satisfaction with a piece of content or genre.
- Real recommendation platforms are, by default, designed to maximize immediate/short-horizon engagement rather than a long-run utility measure — this is asserted about the industry but not independently sourced in the article beyond Knight's own observation.
Inferences:
- Any system that treats an easily-measured, short-run proxy (engagement) as if it were the true, hard-to-measure long-run target (durable satisfaction) is vulnerable to the same self-fulfilling narrowing dynamic — this is a specific instance of Goodhart's Law playing out in consumer recommendation systems, not just in organizational metrics.
A deliberately less-accurate (noisier) recommendation algorithm produces better long-term outcomes for users than a highly accurate one. strength: medium
Explicit premises:
- Moderate prediction error forces the algorithm to occasionally recommend unfamiliar content regardless of current predicted engagement.
- This accidental exposure lets users build consumption capital in styles they'd otherwise never be exposed to long enough to like.
- The benefit is more pronounced when the model includes three or more competing items, since a perfectly accurate system then never gives a new, high-potential item enough exposure to cross the appreciation threshold.
Implicit premises:
- There is some 'moderate' noise level that captures this benefit without recommending so much irrelevant content that it degrades the experience in other ways — the article doesn't state how a real system would find or verify that level.
- Noise and deliberate exploration are being treated as functionally interchangeable solutions, when they are conceptually distinct: noise degrades the algorithm's predictions generally, whereas a targeted exploration policy (e.g., a fixed quota of unfamiliar recommendations) could isolate the same benefit without sacrificing prediction accuracy elsewhere.
Inferences:
- The paper's own logic implies that a platform doesn't need to make its algorithm literally less accurate — it needs a committed exploration policy layered on top of an accurate one, which is a narrower and more actionable design change than 'be worse at predicting.'
Narratives
- Knight's own experience with Spotify: loving several algorithmically-recommended songs at first, then growing sick of the platform repeatedly re-recommending those same 'great find' songs.
→ Grounds the abstract 'straddling' mechanism in lived, concrete platform experience and supplies the personal origin story for the paper. - The hip-hop counterfactual: hip-hop initially sounded 'abrasive and off-putting' to rock-and-roll listeners (as rock once did to earlier audiences); had a Spotify-like algorithm been dominant in the 1980s, its low initial engagement scores might have suppressed the genre's rise entirely.
→ Supports the claim that engagement-optimized algorithms would suppress genuinely valuable but initially unfamiliar genres before an audience has time to learn to like them — and implicitly credits historical human/editorial curation with a genre-incubation role that a pure engagement-algorithm lacks. - Publishing-industry anecdote: the adoption of data-analytics tools reportedly coincided with a rise in trend-chasing among publishing houses, alongside reader complaints that big-publishing-house fiction had come to sound strangely similar.
→ Extends the mechanism beyond music/streaming to books, suggesting the underlying dynamic is general to any content industry that adopts engagement/data-driven curation, not specific to music-streaming algorithms.
Questions addressed
- ✅ Why do well-resourced platforms and publishers, who genuinely want to satisfy their users/readers, end up in 'bad equilibria' where their own curation systems leave people bored or unsatisfied?
Because their algorithms optimize a short-horizon, engagement-based proxy for a preference structure that actually evolves over a much longer horizon, and mistake familiarity-driven engagement for quality — a mechanistic answer formalized in the model. - ✅ Would a deliberately less accurate, noisier recommendation system actually serve users' long-term interests better than a highly accurate one?
Yes, according to the model — noise forces exploration that a perfectly accurate short-horizon optimizer would never undertake on its own. - ⭕ How would a real platform identify or calibrate the 'right' amount of exploration/noise, given the article itself says testing over the necessary multi-year-to-decades horizon is impractical for a technology company to run as an experiment?
The article gestures at a fix ('platforms could... track how often a user has been exposed to specific artistic styles over several years') but doesn't address how such a policy would be tuned, validated, or defended against the same short-term-metric pressures (revenue, immediate churn) that produced engagement-optimization in the first place.
Gaps
- incompleteThe article never states the model's functional forms, parameter values, or sensitivity analysis (how the appreciation curve is shaped, what noise levels were tested, how many Monte Carlo trials produced the reported effect sizes). Without these, a reader can't judge how robust the 'moderate noise helps' conclusion is to the specific modeling choices — a purely theoretical model's conclusions can be highly sensitive to functional-form assumptions.
Read next: Read the underlying paper directly — Samsun Knight, "Engagement-based curation and the evolution of taste," Journal of Cultural Economics (2026), https://doi.org/10.1007/s10824-026-09591-3 — for the actual model specification and robustness checks. - incompleteThe article treats platforms as having essentially one dial (prediction accuracy vs. noise), but real recommendation systems already blend multiple discovery mechanisms — curated new-release rows, human-programmed playlists, explicit 'diversity' injections, cold-start exploration bonuses. It never asks whether real-world systems already implement something functionally like the paper's proposed fix, which would make the normative recommendation partially moot rather than novel.
Read next: Platform-side accounts of recommender-system design (e.g., published descriptions of Spotify's Discover Weekly / exploration mechanisms, or general recommender-systems literature on the exploration-exploitation trade-off) to check how much of this gap is already closed operationally. - uninformedThe study is a purely theoretical, simulated model with no human-subject or real-usage data — the article flags this directly, but doesn't examine whether the inverted-U 'consumption capital' curve (moderate exposure raises liking, excess exposure causes satiation) is itself empirically well-established for genre-level entertainment tastes specifically, as opposed to being carried over from general cultural-economics theory (Stigler & Becker) or adjacent phenomena like the mere-exposure effect and hedonic adaptation.
Read next: Empirical psychology/economics literature on hedonic adaptation and the mere-exposure effect (Zajonc) to check whether an inverted-U specifically holds at the genre/style level for music, film, and fiction — not just for individual songs or images in lab settings. - illogicalThe paper's proposed fix is framed as making the algorithm 'less accurate' or 'noisier,' but the actual causal lever the article describes is forced exploration regardless of predicted engagement — a distinct design choice from genuinely degrading prediction quality. Conflating 'noisy predictions' with 'committed exploration' muddies the practical implication: a platform could keep maximally accurate predictions everywhere and still reserve a fixed fraction of recommendation slots for exploration, capturing the same long-term benefit without the framing cost of 'deliberately recommend worse content.' This is exactly the systems-thinking move the article skips: it treats accuracy and exploration as one dial when the model's own results imply they're two.
Read next: Multi-armed bandit / exploration-exploitation literature (e.g., Sutton & Barto, *Reinforcement Learning*) which cleanly separates prediction accuracy from exploration policy — useful for reframing the paper's own proposed fix as a policy-layer change rather than an accuracy trade-off.