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

AI summary

Overview

Research area: Information retrieval and recommender systems, specifically the economics and content-selection mechanics of free, feed-based content platforms (arXiv:2610.37695v1 [cs.IR], 24 Oct 2026).

Technical level: Beginner-Friendly. The argument is carried by plain-language reasoning, historical analogy, and an explicitly simplified toy model with illustrative parameters; the only formal tools are a basic attention-budget inequality, Bayes' formula, and a Beta-belief update.

Scope: A position paper arguing that AI code-generated "one-click video" floods free platforms with service-type content, that this breaks the attention-funded recommendation pool mechanism, and that a human-curator "BoLe system" is one possible way out.

What This Paper Is About

After Opus 5.5 was released on September 22, 2026, a method spread quickly in which a model writes a program that draws every frame and renders it into a motion-graphics explainer video, producing a wave of "one-click video." The authors argue that this wave is mostly service-type content — watched to meet a need rather than enjoyed for itself — and that because anyone can produce it, its supply grows far faster than work-type content, which is limited by the number of people with something to say. The paper's goal is to show that this imbalance, combined with a fixed total of user attention, causes the layered trial-viewing "recommendation pool" to fail past some growth multiple, and to propose a two-track selection system in response.

Key Contributions

1. A new distinction at the level of the single item: service-type content (watched to meet a need; e.g., most AI short dramas) versus work-type content (watched for the content itself; e.g., a long video by a big YouTube creator, a Netflix series, a Stephen King novel), together with an argument that AI raises the supply of the two kinds by very different amounts.
2. A failure argument for the recommendation pool mechanism: the claim that trial viewing spends user attention, attention is fixed, and therefore selection degrades toward guessing as new content grows — demonstrated with a simplified mathematical model and a sensitivity check.
3. The BoLe system: a proposal that machines and entry rules judge only whether service-type content passes the bar, while trusted people judge work-type content, with machines amplifying their choices.
4. Practical guidance for producers: the claim that the most important thing for content makers now is to know which of the two kinds they make, and to separate the two production pipelines structurally rather than relying on discipline.

Main Findings

• The ratio becomes unbalanced by construction. AI speeds up both kinds of content, but service-type content is open to anyone ("as long as the work is done properly"), while work-type content is capped by the number of creators who want to express something; at the time of writing, AI still cannot decide for the producer what to say.
• The productivity jump is large and personal. A five-minute science video built mainly on motion graphics used to take the authors' team five days (effects made frame by frame in After Effects, then edited in Final Cut Pro); now, with the first author checking each video in person, they can make three to five videos in one day. Dropping professional quality checks and only wanting the video to look good yields twenty videos a day, about one hundred times the past.
• Trial viewing has a hard budget. With c bits of quality information per trial view, εX trial views per day, and N new items, information per item is bounded by cεX/N; if N grows one hundred times, each item gets one hundred times less information.
• Both escape paths are bad. Either the trial-viewing budget stays fixed and viewers per item fall to 1/k of the original (selection becomes like drawing lots), or viewers per item stay fixed and the share of users' time on untested content rises to k times the original — if the baseline is 10%, then at ten times everything users see is untested.
• Diluting good content breaks selection even with accurate trial viewing. Assuming 90% of good content and only 10% of mediocre content advance, about 30% of advancing items are good when good content is 5% of new content, but less than 0.5% are good when it is diluted to 0.05%.
• Toy model results (three layers, 100/500/2500 trial viewers, good content 5% at baseline). The mechanism still has room up to three times growth and the top pool is almost all good content; at ten times, good content in the top pool is only about 12%; at one hundred times, about 0.1%. In the most optimistic case, where good content also grows by the same ratio, one hundred times growth leaves only about 10% good content in the top pool — twice the 5% at input, so selection almost does nothing.
• Sensitivity does not change the direction. Setting each parameter to half and to double its value always lowers top-pool quality as the multiple rises; the multiple at which the good-content share falls below half lies between about three and forty times. Doubling three parameters at once (good-content share, completion-rate gap between good and bad, viewers per layer) pushes that multiple only to between fifty and nearly two hundred times, and implies users spend 20% of their time trial viewing.
• Content made to target the algorithm is a separate failure. Producers who optimize the "first taste" (cover, first three seconds) lower the mutual information I(Q;M) between the metric M and true quality Q; in the toy model, if all added mass-produced content merely passes the trial-viewing metrics, good content in the top pool is only about 7% at just three times growth. Platform subsidies worsen this, since margin approaches the full subsidy as production cost goes to zero.
• Machine pre-review is not sufficient. It can address parts of only three of the platform's seven needs (finding good content, avoiding un-watched storage/distribution/moderation costs, removing traffic cheaters); it cannot tell whether an item expresses something well, cannot keep ad-attracting creators from leaving, and its own per-item cost scales with volume. The trade-off rule the authors describe has platforms protecting proven ad-attracting accounts and narrowing the entrance for newcomers.
• The old production barrier was doing free first-round selection. Before code-generated video, difficulty of production kept most people out, and every platform need quietly depended on that filter; its removal is "the other side of the same thing" as the service-type inflow.
• Revenue share is thinned. TikTok's 2020 US creator fund was a fixed 200 million dollars split by share of views; one well-known creator reported pay falling from about 5 cents to 2.5 cents per thousand views, and the fund was shut down in December 2023. On YouTube, the top 3% of channels took more than 85% of views, and about 96.5% of creators who wanted to earn from ads could not reach the US poverty line.
• The two kinds need different kinds of judgment. Passing the bar is a one-bit question that machines and entry rules can answer; sorting work-type content into K levels needs about log₂K bits, which the authors argue only careful expert judgment supplies — a high-capacity, low-noise channel that does not spend public attention.
• Trust in a BoLe accumulates slowly and conditionally. Starting from Beta(α, β), after m recommendations with h satisfied the belief updates to Beta(α+h, β+m−h), so trust is built one recommendation at a time and cannot be bought outright; this yields a claimed "conditional cumulative advantage" for whichever platform builds a BoLe system first.
• The one-year demo shows a stability threshold. At past production levels the simulated platform stays stable for the whole year (shown at Day 200); once the production multiple goes above three to five times, the pool can no longer pick out good content, users and ad revenue decline, and costs rise with content volume (shown at Day 250 at one hundred times).

Methodology in Plain English

The paper is an argument built from three materials. First, first-hand industry observation: the authors run a team that has made professional short videos for more than ten years, and they describe their own before-and-after production times; they also report that platform numbers for this content have not been published, so the scale of the wave is their own observation as people working in the field. Second, historical analogy: nineteenth-century wet-plate photography, Kodak's 1888 box camera ("You press the button, we do the rest"), the 1900 one-dollar Brownie, and the 1902 Photo-Secession; plus Demand Media and Google's 2011 "Panda" update (about 11.8% of US queries) as a rehearsal of algorithm-targeted mass production. Third, a deliberately simple model: an attention-budget inequality, a Bayes calculation showing how a falling base rate destroys selection, and a three-layer toy simulation with fixed per-layer trial-viewing budgets, followed by one-at-a-time and combined parameter perturbations. A lightly interactive demo then wraps the selection model in illustrative rules for user churn, ad revenue, revenue share, platform costs and creator churn. The authors state repeatedly that all parameters are illustrative, that the conclusion rests on structure rather than specific numbers, and that the platform mechanism description comes from work experience rather than published platform material.

Why This Matters

Impact on research. The paper reframes cold-start and exploration-exploitation (it cites Li et al., 2010 and Wu et al., 2016) as a question of who pays for trial viewing and whether that budget can absorb a supply shock, rather than as an algorithm-design problem alone; it also connects recommender-system practice to information-theoretic limits on what engagement metrics can reveal about quality. It sits alongside Bourdieu's restricted versus large-scale production, Adorno and Horkheimer's culture industry, and Lizzeri's certification economics, while explicitly declining to judge which kind of content has more value.

Real-world applications:

• Platform design: two-track selection in which machines gate service-type content on pass/fail and trusted curators rank work-type content, with algorithms amplifying curator picks.
• Creator strategy: deciding whether to build a service-type pipeline optimized for stability, speed and price, or a work-type pipeline built on expression, and separating the two structurally.
• AI tool and model builders: choosing which kind of content their tooling targets, since that determines what to value and how to measure success.
• Advertiser and buyer decisions: understanding why cheap base supply and curated work command different terms under a fixed attention budget.

Industry relevance. The paper documents that platforms are already tightening entrances: a paid science fiction magazine blocked more than five hundred generated spam submissions in the first twenty days of February 2023 and closed submissions; Amazon's self-publishing platform capped accounts at three new books per day in September 2023; a large programmer Q&A community banned generative-tool answers in December 2022 partly because moderators were overwhelmed. Early BoLe-like forms exist too: Medium's 2023 "Boost" (community nomination, human confirmation, machine distribution), Apple Music's 2015 team of expert editors, and the first author's October 2026 invitation to a closed WeChat Official Accounts test as a BoLe for "Original Picks."

Future Directions

• Test the model against real platform data. The authors concede their mechanism description comes from work experience, not platform disclosure, and that platforms have published no numbers for this content wave; calibrating the toy model and the Figure 2 loop with real data is the obvious next step.
• Does competing BoLe systems beat a monopolistic one? Citing Lizzeri's result that a monopoly certifier reports only pass/fail while competing certifiers give detailed quality information, the paper suggests BoLe systems may themselves be monopolistic or competitive, and leaves that unresolved.
• What does a following actually mean? When readers follow a BoLe, do they follow because he chooses well (making the BoLe a service), or because of the person himself (giving the BoLe's judgment a work-type side)?
• Scale of compliance and moderation. With the paper's one-hundred-times scenario, review work also grows one hundred times, while machine pre-review costs money per item; how that cost is borne is left open.

Target Audience

Content producers and studios deciding how to organize production, platform and recommender-system practitioners responsible for cold-start and content supply, builders of AI creative tools and models, and researchers in information retrieval or media economics interested in how generative supply shocks interact with attention-limited selection. It is written to be read without a technical background: the argument is conceptual and the mathematics is elementary.

Authors’ abstract

In September 2026, a new large language model was released. 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. "One-click video" content then appeared in large amounts on many platforms. On the surface, this is a rise in creative power at the level of the individual. We argue that, at a deeper level, it changes the structure of the content industry. This judgment rests on two observations. First, this wave of content is service-type content. People watch it to meet a need, not to enjoy the content itself. The other kind is work-type content. AI makes both kinds faster to produce. But anyone can come in and make service-type content, while work-type content is limited by the number of people who have something to say. So the ratio between the two slowly becomes unbalanced. Second, free platforms choose content through layers of trial viewing, and this uses up user attention. The total amount of attention is fixed. The unbalanced ratio also dilutes good content. A simplified mathematical model shows that, as long as these conditions hold, the recommendation pool mechanism will fail once new content grows past a certain multiple. Based on this, we discuss one possible way out. Machines and entry rules judge whether service-type content passes the bar. Trusted people judge work-type content. We call this the BoLe system. At the end, we briefly imagine how the two kinds of content may develop from here.