Scale over Preference: The Impact of AI-Generated Content on Online Content Ecology

Abstract

The rapid proliferation of Artificial Intelligence-Generated Content (AIGC) is fundamentally restructuring online content ecologies, necessitating a rigorous examination of its behavioral and distributional implications. Leveraging a comprehensive longitudinal dataset comprising tens of millions of users from a leading Chinese video-sharing platform, this study elucidated the distinct creation and consumption behaviors characterizing AIGC versus Human-Generated Content (HGC). We identified a prevalent scale-over-preference dynamic, wherein AIGC creators achieve aggregate engagement comparable to HGC creators through high-volume production, despite a marked consumer preference for HGC. Deeper analysis uncovered the ability of the algorithmic content distribution mechanism in moderating these competing interests regarding AIGC. These findings advocated for the implementation of AIGC-sensitive distribution algorithms and precise governance frameworks to ensure the long-term health of the online content platforms.

Figures (3)

• Figure 1: The scale-over-preference dynamic of AIGC in content creation and consumption.(a) Integration of AIGC into the studied platform. The studied platform has invested heavily in generative AI, with daily expenditures on AIGC reaching 4 million USD across the entire platform Kuaishou2025Q1. Specifically within its Local Life channel, about 35% of creators have adopted AIGC tools, and AIGC videos account for 30% of total uploads on average.(b) Conceptual framework. Schematic illustration of how AIGC reshapes the content ecology through changes in creator production, consumer engagement, and algorithmic distribution. (c) Creator production and engagement returns. Complementary cumulative distribution function (CCDF) plots for matched creators ($n = 2,497$ pairs) show that AIGC creators exhibit substantially higher productivity (rightward shift in video volume; left panel) while achieving aggregate engagement returns—measured by total valid views (middle) and full views (right)—comparable to those of HGC creators. Valid views denote views exceeding the platform duration threshold, and full views denote videos watched to completion. (d) Consumer engagement preference. Matched interaction pairs ($n = 47,288$pairs) reveal systematically lower engagement with AIGC than HGC. AIGC interactions yield lower mean valid-view and full-view rates (left) and shorter view duration (right). Left subpanels report means with 95% CIs; right boxplots show median (line), mean (triangle), IQR (box), and whiskers at $1.5\times$ IQR. (e) Scale-over-Preference index (SoPI) for each day. Daily observations ($n = 61$ days) map relative AIGC supply scale ($S$) against relative consumer preference on AIGC ($P$), colored by SoP index $SoPI =ln(S/P)$. Most days fall in the high-supply/low-preference region above the $SoPI=ln(1.5)$ contour (dashed), indicating a persistent divergence between production scale and observed preference.
• Figure 1: Decomposition of creation volume and engagement returns for AIGC creators in creator-side comparison. The complementary cumulative distribution functions (CCDFs) compare the production and engagement metrics between HGC creators (solid blue line) and AIGC creators (solid green line). To isolate the source of AIGC creators' productivity and engagement, their metrics are further decomposed into their AIGC video contributions (dashed green line) and HGC video contributions (dashed blue line). The panels illustrate (left) Creation Volume, (middle) Valid-View Count, and (right) Full-View Count.
• Figure 2: Algorithmic content distribution mechanisms moderate the scale-over-preference dynamic. (a) Algorithmic exposure disadvantage and compressed lifecycle of AIGC. CCDFs for matched video pairs ($n = 178,854$ pairs) show that AIGC receives lower cumulative exposure than HGC (left) and exhibits a more compressed exposure lifecycle (right; days to reach 90% of 31-day exposure), consistent with weaker engagement preference for AIGC. (b) Algorithmic exposure response to AIGC supply expansion. Granger analysis (left) with all negative lag coefficients indicates that increases in daily AIGC supply precede reductions in the median exposure of newly uploaded AIGC ($p = 0.022$; joint F-test). Regressions across exposure tiers (right) show that higher supply is associated with a downward shift in the exposure distribution: a positive association in the lowest tier ($\leq$10 views) and negative or insignificant associations in higher tiers (11–100, 101–1000, $>$1000 views). (c) Algorithmic exposure response under varying SoP dynamics. Dynamic regressions quantify exposure responses to rising tension between scale and preference. Conditional on relative preference for AIGC, AIGC exposure declines with increasing relative supply ($\beta$ = $-$0.936, 95% CI = [$-$1.089, $-$0.794]; left). Conditional on AIGC relative supply, AIGC exposure decreases with lower relative preference ($\beta$ = 2.683, 95% CI = [1.618, 3.696]; right). Both indicate the algorithm's moderation of AIGC visibility as SoP intensifies. (d) Creator outcomes under algorithmic adjustment. Log-log elasticities show that the overall creator engagement return elasticity with respect to AIGC supply is substantially smaller than for HGC, indicating smaller marginal returns to AIGC expansion ($SOPI$$\uparrow$). Bars denote coefficients and error bars 95%. (e) Consumer outcomes under algorithmic adjustment. Log-log regression elasticities of overall consumer engagement depth remain near zero as AIGC supply increases, indicating that overall consumer engagement does not decline under algorithmic moderation of AIGC expansion ($SOPI$$\uparrow$). Bars denote coefficients and error bars represent 95% CIs, and HGC results are reported for reference. (f) Heterogeneous responses across algorithms. CCDFs of AIGC exposure ratios under two mechanisms show a leftward shift for the population feedback-driven algorithm relative to the individual feedback-driven algorithm, indicating that algorithm design differentially moderates the SoP dynamic.