AI-made videos match humans in total attention by flooding the feed: AIGC creators win aggregate engagement through sheer volume even though viewers prefer human-made clips, and the platform’s recommendation algorithm largely determines how much AI content viewers actually see — suggesting platforms should adopt AIGC-aware distribution and governance rules.

Main Finding

The paper documents a persistent "scale-over-preference" (SoP) dynamic for AI-generated content (AIGC) on a large Chinese short-video platform: AIGC creators produce much more content, and through scale achieve aggregate engagement comparable to human-generated content (HGC) creators, despite systematic consumer preference for HGC. Platform recommendation algorithms, however, actively moderate this tension by down-weighting AIGC exposure as AIGC supply grows, producing smaller marginal returns to additional AIGC and preserving overall consumer engagement depth.

Key Points

• Scale-over-Preference (SoP) pattern

  • AIGC creators are substantially more productive (higher upload volume) than matched HGC creators.
  • Consumers show weaker engagement with AIGC than HGC: lower valid-view and full-view rates and shorter view durations (e.g., valid-view rate mean difference ≈ −0.076, p < 0.001).
  • Despite weaker per-item preference, aggregate engagement of AIGC creators ≈ HGC creators because of higher output volume.
  • The SoP index SoPI = ln(S/P) (S = relative AIGC supply; P = relative consumer preference) is persistently positive across the observation period, indicating higher supply relative to preference (typical S ≈ 0.55–0.70, P ≈ 0.30–0.40).

• Algorithmic moderation of AIGC visibility

  • Matched-video analyses (n ≈ 178,854 pairs) show AIGC receives lower cumulative exposure (show counts) and a more compressed exposure lifecycle (reach 90% of exposure ≈ 2 days earlier).
  • Time-series / Granger tests on daily aggregated data (June 2024–May 2025) indicate increases in AIGC supply Granger-cause reductions in median exposure of new AIGC (joint F-test p = 0.022).
  • Exposure distribution shifts with supply expansion: higher AIGC supply increases share of very-low-exposure items (≤10 shows) and reduces representation in higher-exposure tiers.
  • Regressions conditional on supply/preference: AIGC exposure decreases with higher relative supply (β ≈ −0.936, 95% CI [−1.089, −0.794]) and increases with higher relative preference (β ≈ 2.683, 95% CI [1.618, 3.696]).
  • Elasticities: creator engagement return elasticity wrt AIGC supply is substantially smaller than HGC (smaller marginal returns); consumer engagement depth (valid/full-view rates) remains approximately stable (near-zero elasticities) as AIGC supply grows.

• Heterogeneity by algorithm design

  • Population-feedback-driven recommendation suppresses AIGC visibility more strongly than individual-feedback-driven recommendation (left-shifted CCDF of AIGC exposure ratios), implying algorithm design choices materially affect SoP outcomes.

Data & Methods

• Platform/context
  • Data from the Local Life channel of a leading Chinese short-video platform (platform-wide >400M daily active users). By mid-2025, platform-reported adoption: ~35% of active creators used the platform's AIGC tools in this channel, and AIGC ≈ 30% of uploads in the channel; platform reported large AIGC investment (cited $4M/day across platform).
• Data span and scope
  • Full dataset: June 2024–May 2025 (≈118 million videos for dynamic analyses).
  • Focused behavioral window for creator/consumer analyses: April–May 2025.
  • Samples used in key analyses: matched creators (n = 2,497 pairs for one reported CCDF), broader creator sets (10,763 AIGC creators; 376,713 HGC creators per classification rules), matched user-video interaction pairs (n = 47,288 pairs), matched video pairs for exposure (n ≈ 178,854 pairs), and a balanced sample of 0.8M AIGC vs 0.8M HGC videos for other consumption analyses.
• AIGC identification
  • AIGC labeled via platform-generated metadata tags for content created with official AIGC tools; external-AI usage in unlabeled set assessed with Sightengine and reported as minimal (additional validation in SI).
• Metrics
  • Creator-side: upload volume; aggregate valid views and full views per creator (valid view = watch beyond platform duration threshold; full view = watched to completion).
  • Consumer-side: valid-view rate, full-view rate, view duration per interaction.
  • Distributional: exposure = show count in a 31-day post-upload window; days to reach 90% of 31-day exposure.
  • SoPI = ln(S/P) where S = relative supply scale of AIGC to HGC and P = relative consumer preference for AIGC to HGC.
• Identification/causal tools & robustness
  • Matching procedures to reduce confounding (creator and video matching on characteristics and categories).
  • Granger-causality tests for temporal precedence between supply and exposure changes.
  • Log-log regressions to estimate elasticities; dynamic regressions to assess joint effects of supply and preference.
  • Multiple robustness checks referenced in Supplementary Information (SI).

Implications for AI Economics

• Supply-side incentives and attention allocation
  • Generative-AI lowers marginal cost of production, prompting high-volume AIGC supply that can outcompete HGC in aggregate visibility unless algorithms counterbalance supply growth. This produces a supply-driven attention allocation regime that can reward scale over quality/preference alignment.
• Platform governance and incentive design
  • Algorithms materially shape market outcomes: platform-level controls (exposure adjustments, tiering, feedback-aggregation design) can blunt or amplify SoP dynamics. Designers can tune exposure rules to protect consumer experience and content diversity.
  • Policy levers include differential exposure weighting, caps on automated production, quality signals that prioritize preference-aligned content, and transparency/reporting for AIGC prevalence.
• Welfare and market structure
  • Short-term: algorithmic moderation can preserve consumer engagement metrics even as AIGC supply grows.
  • Long-term risks: persistent scale-over-preference could incentivize low-quality mass production, degrade content diversity, and shift creator returns towards those using automated production — potentially creating winner-take-most dynamics or reducing incentives for higher-quality human creation.
• Measurement and regulatory monitoring
  • The SoPI construct offers a quantitative monitor for platform regulators and firms to track misalignment between supply and preference and trigger governance responses.
• Design trade-offs and heterogeneity
  • Different recommendation architectures (population vs individual feedback) produce different externalities; economic evaluation of algorithm designs should account for their distributional effects on creators, consumers, and platform health.
• Research and policy directions
  • Need for longitudinal welfare analyses: how AIGC-induced supply shifts affect creator labor markets, content diversity, and long-run user retention.
  • Consider experiments/policies (e.g., exposure quotas, differential pricing of recommendations, quality-weighted rewards) to realign incentives and prevent arms races of automated production.

Caveats: observational study on a single platform/channel; reliance on platform AIGC labels (validated but not infallible); algorithm internals are treated as black-box responses inferred from exposure patterns rather than via access to model parameters. The authors report multiple robustness checks in SI.