After One-Click Video: The Inflow of Service-Type Content, the Failure of the Recommendation Pool Mechanism, and the BoLe System
Bo Zou and Chao Xu argue that AI-generated video is not merely making individual creators more productive. In their view, it is changing the economics of free, feed-based content platforms by sharply increasing the supply of a particular kind of material: content made to satisfy a practical need rather than to be valued as a work in itself.
The paper calls this distinction service-type content and work-type content. A short explainer that answers a question, an AI-generated drama, or a video optimized to deliver information quickly belongs to the first group. A documentary, novel, or creator-led investigation that viewers follow because of the creator’s perspective belongs to the second. The distinction is about why people watch, not about length, cost, or production quality.
AI expands the two categories unevenly
The authors say AI improves production efficiency for both categories. They cite the first author’s team, which reportedly moved from producing one motion-graphics science video in roughly five days to producing three to five in a day with human checking and responsibility retained. That example is a single team’s experience, not an industry-wide measurement.
The important difference, in the authors’ model, is each category’s potential supply. Anyone can enter service production if the goal is to meet a defined need and clear a minimum quality bar. Work-type content still depends on people who have something distinctive to express. Better tools can help those creators execute faster, but they do not automatically create more original viewpoints.
The result, the paper argues, is a growing imbalance: service-type content can expand far faster than work-type content. The authors compare this with ready-made food or mass-market photography. Lowering the production barrier increases access and convenience, but it also removes a layer of selection that once kept supply constrained.
Why recommendation pools become fragile
Most free platforms use staged exposure to evaluate new material. A post or video is shown to a small audience, and items with strong early signals are promoted to larger pools. This approach addresses the cold-start problem while using viewer behavior as a relatively inexpensive quality signal.
But every trial view consumes attention. The authors formalize that constraint with a simplified model: if the number of new items rises while the available trial-viewing budget remains fixed, each item receives less evidence. If platforms preserve the number of trial viewers per item instead, a larger share of users’ time is spent testing content rather than consuming content that has already earned distribution.
The paper’s illustrative model uses three recommendation layers and hypothetical parameters. It reports that the system has room when supply grows modestly, but that the share of good content in the top pool falls rapidly as the growth multiple increases. Under one set of assumptions, content volume growing tenfold leaves about 12% good content in the top pool, while a hundredfold increase leaves about 0.1%. These are model outputs, not measurements of a real platform.
The authors also highlight a base-rate problem. Even an accurate test can produce mostly mediocre results when genuinely good items become rare. Producers may further exploit the system by optimizing covers, opening seconds, completion rates, or other measurable signals. That can make the metrics less informative about the underlying quality. In the paper’s analogy, recommendation systems risk selecting content with an appealing surface but poor substance.
The proposed BoLe system
The paper argues that machine screening alone cannot solve the problem. Automated review can check whether service-type content meets a defined threshold, detect some low-quality or abusive material, and help manage scale. It cannot reliably determine whether a work-type item is meaningful, original, or worth sustained attention.
The proposed alternative is the BoLe system, named after the traditional Chinese evaluator associated with recognizing valuable horses. Machines and entry rules handle pass-or-fail judgments for service content. Trusted experts or curators evaluate work-type content, after which algorithms amplify their selections.
The authors point to existing arrangements that resemble parts of this system, including human nomination and confirmation mechanisms on publishing platforms and expert-curated music playlists. Their central claim is that a careful expert judgment can provide more informative evidence about a work than a quick user interaction, without consuming the attention of the entire audience.
Trust, however, is the difficult part. The paper models trust as something updated through repeated recommendations and reader satisfaction. A curator may be appointed quickly, but credibility accumulates gradually. The authors therefore speculate that platforms able to recruit respected curators early could gain a conditional advantage, while emphasizing that network effects, user habits, and capital would also shape outcomes.
Practical implications and limits
For creators, the paper’s most actionable advice is to decide which product they are making. Service-type content should be designed around reliability, speed, useful coverage, and a clear quality threshold. Work-type content should protect judgment, voice, and the reason a particular creator deserves attention. The authors describe splitting these functions into separate production pipelines within their own team, but present that experience only as an illustration.
For platforms, the argument suggests that recommendation metrics cannot scale indefinitely when content supply grows much faster than user attention. More sophisticated ranking may improve allocation, but it cannot create additional viewing time. Platforms may need clearer quality gates, expert programs, trusted channels, or business models less dependent on advertising and undifferentiated feed competition.
The authors are explicit about the evidence limits. Platform data on the claimed one-click-video wave were not available, and the timing and scale are based on industry reports and the authors’ observations. The recommendation model uses illustrative parameters, while the broader platform simulation includes uncalibrated rules for churn, advertising, costs, and creator revenue. The paper offers a structural hypothesis and a possible institutional response, not a validated forecast.
Its lasting contribution may therefore be conceptual rather than predictive. As AI makes adequate content nearly free to produce, discovery becomes less about finding enough material and more about deciding what deserves scarce human attention. The BoLe system is one proposed answer: automate the threshold, but reserve high-information judgment for people whose recommendations have earned trust.
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.