The Curatorial Bottleneck: Why Selection Cannot Scale Like Production

Every June, in a five-story Georgian building on Piccadilly, an event takes place that has occurred annually since 1769 and that the British art world treats with the gravity of weather. The Royal Academy Summer Exhibition is the oldest continuously running open submission exhibition in the world. Any artist, anywhere on the planet, can submit work. In 2024 and 2025, roughly eighteen thousand pieces arrived. The selection panel — composed of Royal Academicians serving short rotating terms — accepted approximately two thousand. The ratio is around 11%. The work they had to look at would, if printed at A2 and laid in a single line, stretch the length of central London.

The selection happens over a series of days called the Hangs. Royal Academicians stand in front of a moving conveyor and have, in the Academy’s own publicly described practice, somewhere between a few seconds and a minute to make each decision. Yes or no, in or out, hung or rejected. They are looking at the lifetime output of strangers and deciding the fate of each piece in less time than it takes to read this paragraph. They are, by any reasonable standard, the bottleneck — and they have always been the bottleneck, since the year before the United States existed, because the number of artists in Britain who want to be exhibited has never been smaller than the number of Academicians available to look at their work.

This essay is about that ratio, and about every place it recurs, and about why the gap between production capacity and selection capacity keeps widening — and what the people running the systems most affected by it can actually do.

A Law That Bites

The asymmetry is mechanical, not cultural. Production scales with effort, money, automation, and now with generative models that have driven the marginal cost of plausible content close to zero. Selection scales with attention, and human attention has not changed measurably in capacity in any of those time horizons.

A photographer in 1869 could expose three plates an hour. A photographer in 2025 with a phone produces several thousand passable images in the same time. The number of editors who can look at those images and decide which are worth showing has not increased to match. It cannot.

This is a law in the sense that it is structural. You can build a faster camera. You cannot build a faster eye. You can build a model that drafts a thousand essays a day. You cannot build a reader who can absorb a thousand essays a day. The two sides of the equation have different floors and the production side has been falling toward zero, dragging the curation side into the role of bottleneck for every system that produces more than its readers can keep up with.

In a 2024 study from Asana’s Work Innovation Lab, researchers found that 65% of knowledge workers reported that AI tools created more coordination work for them than they removed, with the figure rising to roughly 90% among self-identified high-AI-usage workers. The result has been widely cited under the label “the productivity paradox of AI.” The paradox is only paradoxical if you forget about the curatorial side of the ledger. AI does not actually create more work; it creates more output, and someone has to look at the output, decide what to do with it, route it, integrate it, edit it. That someone is a human. That someone is the bottleneck. The faster production goes, the more pronounced the bottleneck becomes.

This is what the music industry stopped trying to deny around 2010, when Spotify’s catalog crossed about 20 million tracks and the human-curated radio model became visibly hopeless. Spotify launched Discover Weekly in July 2015 as an algorithmic answer to the curation problem — collaborative filtering plus content-based features producing a personalized 30-track playlist every Monday. The product was, and is, genuinely good. But notice what it did: it converted the curation problem from a human selection problem to an algorithmic selection problem. The bottleneck did not disappear. It moved.

It is worth being precise about what algorithmic curation actually does. It uses prior human selections as input. Spotify’s recommendations work because millions of humans have made millions of selection decisions that the algorithm can interpolate. The algorithm is not curation; it is amortized curation. The actual judgment was performed earlier, by people, and the algorithm is doing pattern-matching over their fossilized choices. When the upstream production rate exceeds the rate at which humans are seeding the algorithm with new selection decisions, the algorithm runs out of fuel and starts to repeat itself, recommend the same artists to everyone, and visibly homogenize.

The AI Slop Economy

The current acute version of the problem has acquired a name. It is being called “AI slop” — the flood of mediocre AI-generated content arriving in feeds, marketplaces, news sites, app stores, Amazon, Etsy, Spotify, every platform where humans submit and the platform displays. In November 2023, the journalism site Futurism reported that Sports Illustrated had been publishing articles under fake AI-generated bylines accompanied by AI-generated headshots, a story that broke open a wider conversation about the indistinguishability problem. Around the same time, the Wikipedia talk pages began to fill with discussions of AI-edit detection and policy enforcement, because volunteer editors were beginning to feel they were spending more time triaging suspected machine-generated edits than improving articles.

The economic frame on this is sharp. When production cost approaches zero, the only scarce input is selection. The value of the supply chain concentrates at the curation step, because every other step is now infinitely available. Whoever decides what gets in front of human eyes is the one capturing the value, because that decision is the one that humans cannot delegate without losing something they wanted.

Curation requires judgment. Judgment requires the curator to be willing and able to refuse the bulk of what is submitted. That willingness is structurally fragile under economic pressure: if the curator’s revenue depends on the submitters, the curator’s “no” gets soft, and the gate widens, and the floodwater that was being held back rushes through. Curation under economic pressure decays into rubber-stamping, and rubber-stamping is not curation.

The Royal Academy can be hard about the 11% acceptance rate because the institution does not depend on submission fees for survival. Most curators are not in that position. Most magazines depend on the very advertisers whose products their editorial judgments might inconvenience. Most app stores depend on commission from the apps whose quality they are notionally screening. Most social platforms depend on engagement metrics generated by exactly the content their content-moderation teams are notionally restraining. The economics of curation are at war with the act of curation, and the war is being lost in slow motion almost everywhere.

What Curators Actually Do

To understand why selection cannot scale the way production can, it is worth being more precise about what the act of curation actually consists of. The Royal Academician glancing at a painting for ten seconds is not doing nothing. They are bringing two decades of trained attention to a moment of decision. The same model Jonathan Sterne worked out for medical auscultation in his 2003 book The Audible Past applies here: trained perception is not a faster version of untrained perception; it is a categorically different operation. The Academician sees compositional choices, technical execution, conversation with art history, the question of whether the work has anything to say. A first-year art student cannot do this. The capacity is built across thousands of looking-hours, and the resulting ten-second judgment is the compressed output of those hours.

This is true in every domain that has serious curators.

All of them spend seconds-to-minutes per submission and reject the overwhelming majority, and they are right to do it that way because what they are bringing to the decision is not the seconds — it is the accumulated context that lets those seconds be diagnostic. The ten seconds is the visible part of a process that took two decades of preparation.

What this means, practically, is that you cannot scale curation by hiring more curators. You can hire more bodies. You cannot manufacture trained attention. The pipeline that produces a senior curator is not a pipeline that admits to acceleration. It involves apprenticeship, exposure to enormous volumes of high-quality and low-quality work, mentorship from previous senior curators, and time. Several years of time, at minimum. Often a decade.

Where Algorithmic Curation Helps and Where It Fails

Algorithmic curation does real work and saying otherwise is dishonest. Spotify’s recommendation engine has introduced a generation of listeners to artists they would not have found through human radio programmers. Google Search, for all its problems in 2026, still surfaces useful pages from a corpus of billions in milliseconds. Amazon’s product recommendations are good enough that a small fraction of total purchases originate from them, which at Amazon’s scale is many billions of dollars of value.

What algorithmic curation cannot do, and is increasingly seen to be unable to do, is bring trained judgment to the question of whether something should exist. Algorithmic curators score what already exists; they do not gatekeep what enters. When a content marketplace flips from human-curated to algorithm-curated, the gatekeeping is gone. Everything gets in. Everything competes for the algorithmic ranking on the same axes. The axes the algorithm can score are surface-level (engagement, click-through, dwell time, similarity to liked things). The axes it cannot score are deep (originality, courage, whether the work has anything to say). The selection function therefore drifts toward optimizing the surface and ignoring the deep, and the result, in practice, is the lo-fi soundscape complained about in every contemporary critique of recommendation feeds.

The honest characterization of what algorithmic curation provides: it is competitive ranking among items that have already cleared a gate. When the gate is wide open, ranking is a poor substitute for selection, because the worst item ranked first is still worse than the best item that did not exist. The job of selection is to decide what enters the ranking pool at all, and algorithms have, so far, been unable to do that job at the quality level humans bring.

What to Do on Monday

If you run a system that produces faster than your team can review — and most teams in 2026 do — there are three habits to install.

First, measure the gap explicitly. Count, for one week, the number of artifacts your system produces and the number of artifacts your team actually evaluates. If the ratio is more than 5:1, you have an unmanaged curation problem and your unevaluated output is functionally noise. Most teams do not measure this number and so cannot notice when it gets worse, which it is doing.

Second, invest in your curators before you invest in more production capacity. The marginal value of a faster producer is low when the bottleneck is downstream. The marginal value of a faster curator is high, and the marginal value of a better-trained curator is higher still. This means apprenticeship time, exposure to representative work, deliberate practice in making and reviewing selection decisions. The honest investment horizon is years, which is uncomfortable to admit in a quarterly-OKR culture, but the math is the math.

Third, protect the economic independence of your selection function. The single fastest way to break a curatorial system is to make the curator’s compensation depend on the people they are supposed to be filtering. Where you can structurally insulate the gatekeeper from the gated — through endowments, through subscription revenue, through institutional funding — the gate stays open in the right direction and closed in the other. Where you cannot, watch the acceptance rate over time and you will see the bottleneck dissolving into rubber-stamp territory.

The Royal Academy has been managing this same problem for two hundred and fifty-six years. They have settled on a low acceptance ratio, a rotating selection panel composed of practitioners rather than administrators, an institutional structure that insulates the panel from the submitters, and a public ritual — the Summer Exhibition — that makes the curation legible as a result. Most contemporary digital platforms do none of these things. They have low or zero gatekeeping, algorithmic ranking of what gets through, an economic model that ties the platform’s revenue to submitter behavior, and no public ritual that holds the curator accountable for the quality of what was selected. The result, predictably, is what we have.

The next decade of platform design will be defined by the people who take the curatorial bottleneck seriously and build economic structures around it that do not collapse the moment production gets faster. Those structures cost more than algorithmic curation. They scale worse. They cannot match production velocity. That is the point. The job they do is the job that production velocity cannot do for itself. The faster production gets, the more valuable the slow human standing at the gate becomes, because they are the only thing left between the firehose and the audience.

If you want to know what the most valuable role in your organization will be in five years, it is the person who can say no to most of the work and be trusted that their no is right. That person is not faster than the firehose. They do not need to be. The firehose does not know what it is producing. They do.

Sources: Royal Academy Summer Exhibition public reporting on submission counts and acceptance rates (2024–2025); Asana Work Innovation Lab 2024 study on AI-induced coordination work; Spotify Discover Weekly launch (July 2015); Futurism reporting on Sports Illustrated AI-generated bylines (November 2023); Wikipedia talk-page discussions on AI-edit detection (2023—ongoing); Jonathan Sterne, The Audible Past (Duke University Press, 2003); editorial-independence essays at Smarter Articles and other 2025 publishing; Royal Academy practice on the Hangs as publicly described by the institution. Herbert Simon’s 1971 quote on attention as the scarce resource of the information age is the philosophical anchor for the production-vs-selection asymmetry.