After One-Click Video: The Inflow of Service-Type Content, the Failure of the Recommendation Pool Mechanism, and the BoLe System

After One-Click Video. The Inflow of Service-Type Content, the Failure of the Recommendation Pool Mechanism, and the BoLe System. By Bo Zou and one other. Published on the twenty-fourth of October.

The starting point is a recent event. In September two thousand and twenty-six, a new large language model was released, and soon after a new way of making explainer videos became very popular. People asked the model to write a program, and the program produced the video. This one-click video content then appeared in large amounts on many platforms.

Read at the surface, that is a rise in creative power for individuals. The authors argue for a deeper reading: that it changes the structure of the content industry itself, and that the mechanism free platforms use to pick content, the recommendation pool, will fail under it once new content grows past a certain multiple.

The argument rests on two observations. The first is about what kind of content this is. The authors split content into two kinds. Service-type content is watched to meet a need, not for the enjoyment of the content itself. Work-type content is the other kind. The one-click wave, they say, is service-type.

AI makes both kinds faster to produce, but the two are limited differently. Anyone can come in and make service-type content. Work-type content is limited by the number of people who have something to say. Over time, then, the ratio between the two slowly becomes unbalanced.

The second observation is about how free platforms choose what to show. They do it through layers of trial viewing, and each layer uses up user attention. The total amount of attention is fixed. So a growing flood of new content draws on a budget that does not grow, and the unbalanced ratio dilutes the good content inside it as well.

These conditions are put into a simplified mathematical model. The model shows that, as long as the conditions hold, the recommendation pool mechanism fails once new content passes a certain multiple.

The result is conditional, and stated that way. The failure follows only as long as those conditions hold: attention is fixed in total, layers of trial viewing consume it, and the ratio of service-type to work-type content keeps drifting out of balance. It applies to free platforms that choose content in this layered way. And the model behind it is, in the authors' own word, a simplified one, so the claim is about the structure of the mechanism under those assumptions.

What follows from the model is offered as one possible way out, not the only one. The authors call it the BoLe system. It splits the judging along the same line as the content. Machines and entry rules judge whether service-type content passes the bar. Trusted people judge work-type content. The paper closes by briefly imagining how the two kinds of content may develop from here.

The core move is a distinction. Once content is sorted by why people watch it, the one-click wave stops looking like more of everything and starts looking like a lopsided supply: one kind that anyone can produce, and one kind bounded by people with something to say. A selection system that spends fixed attention testing every new item is then, on this account, a system with a breaking point, and the BoLe proposal is to judge each kind by a different gatekeeper.

After One-Click Video. The Inflow of Service-Type Content, the Failure of the Recommendation Pool Mechanism, and the BoLe System. By Bo Zou and one other. Published on the twenty-fourth of October. This is the long version: the whole paper rather than the abstract.

After Opus 5.5 was released in September, a new method spread fast: ask the model to write a program that draws every frame, then render it into a polished motion graphics explainer. Soon one-click video tutorials were everywhere.

The first author's team has made professional short videos for more than ten years. A five-minute science video used to take five days. Now, with the author checking each one in person, the team makes three to five a day. Drop the quality checks, and it becomes twenty a day: about one hundred times the past. The question is what happens to a free, feed-based platform when that multiple arrives. The answer offered is that the machinery for choosing what you see breaks.

Two ideas come first. Service-type content is watched to meet a need, like most AI short dramas and most of this explainer wave. Work-type content is watched to enjoy the thing itself: a Netflix series, a Stephen King novel. The difference is not cost or length. AI speeds up both, but anyone can make service-type content, while work-type content is limited by the number of people with something to say, and better tools do not multiply them. So the ratio slowly tips toward service, much as snapshots flooded daily life after Kodak's box camera.

The second idea is the recommendation pool. A new item is first shown to roughly a hundred people. If it beats similar items, it moves up to roughly five hundred, and is amplified layer by layer. User feedback is the most accurate and cheapest judgment available.

The argument rests on a premise rarely spoken aloud: trial viewing spends user attention, and the total amount of attention is fixed. Each trial view trades a few seconds of someone's time for a little information about quality. Multiply the information per view by the trial views users can spare in a day, divide by the number of new items, and you have the most information any item can receive. Only the bottom of that fraction grows. A hundredfold more items means each gets a hundred times less evidence, until judgment is close to a guess.

So the platform has two paths. Keep the trial budget fixed, and each item's audience shrinks until completion rates are mostly noise and selection becomes drawing lots. Keep each item's audience fixed, and the share of users' time spent on untested content rises by the same multiple. Starting from ten percent, at ten times everything users see is untested. Any compromise sits between two bad ends, and dilution then works on whatever survives.

Two more turns. Producers chasing traffic target what trials can measure: the cover and the first three seconds. The paper's image is chocolate-flavored poop, which passes the entrance and gets amplified, against poop-flavored chocolate, good but unpleasant at the entrance, which gets kept out. In information terms, optimizing for the metric lowers the mutual information between the metric and true quality. And machine pre-review is not enough: it can block garbage below the bar, but not the large amount of service-type content above it.

The evidence is a toy model with three trial layers, every parameter illustrative. Good content starts at five percent. The layers use a hundred, then five hundred, then two thousand five hundred viewers per item, each layer's total budget fixed. Up to about three times the old volume, the top pool stays almost all good. Then it drops fast: about twelve percent good at ten times, about a tenth of a percent at a hundred times.

The comparison that most directly tests the thesis is the optimistic one. Let good content grow in proportion too. At a hundred times, the top pool is still only about ten percent good, twice what went in. Selection is barely working, with no dilution at all. And if the added mass content just passes the trial metrics, good content in the top pool falls to about seven percent at only three times.

Sensitivity runs set each parameter to half and to double. One at a time, the multiple where good content falls below half lands between about three and forty times. Doubling three platform-friendly parameters together pushes it to between fifty and nearly two hundred times, but then users spend twenty percent of their time on trial viewing.

There is a Discussion and limitations section, and it is candid. The model's parameters are illustrative, and the authors claim the conclusion depends on structure, not numbers. How platforms run trial viewing comes from the authors' work experience, not from published material. The scale of the video wave rests on media reports and the authors' observation, since platforms have released no data. The team's case is one sample. A year-long demo adding user, ad and creator churn uses uncalibrated rules, and its numbers are explicitly not evidence. The authors call the outlook one that may be wrong.

So nothing here is measured on a real platform, and the proposed way out is argued with examples, never evaluated.

That way out is named after Bo Le, a man in ancient China known for judging horses. If attention is the only source of information about quality, a second source is needed that does not spend the public's attention. For service-type content, passing the bar is a one-bit question that machines and entry rules can answer. Ranking work-type content needs far more bits, and a trusted expert watching carefully is a channel with large capacity and low noise. People choose, machines amplify. Medium's Boost is an early form. Trust in such curators builds one recommendation at a time, so money cannot buy it all at once. And for producers, first know which kind of content you make.