Title: Incentivizing Engagement in the Kidfluencer Landscape: A Multimodal Weak Supervision Framework

Abstract

The proliferation of child content creators, or "kidfluencers," on YouTube has sparked significant ethical debate regarding the exploitation and digital labor of minors. Although new legislative efforts aim to govern this sector, empirical data connecting exploitation to audience engagement remains limited due to the challenges of measuring exploitation at scale. To address this gap, this research employs a multimodal AI audit involving 5,051 videos from 79 kidfluencer channels. Utilizing weak supervision, the study identifies exploitation indicators without relying on extensive manual labeling.

The methodology aggregates noisy labeling functions, such as LLM-driven title classification and GPT-4 Vision analysis of thumbnails and video descriptions, across six dimensions established in existing literature. These inputs generate a probabilistic exploitation score for each video. A validation study involving 107 annotators demonstrated robust alignment with human assessment, achieving a macro-average F1 score of 0.911. Furthermore, the model exhibited high sensitivity in detecting overall exploitation risk, with a recall rate of 0.960 and an F1 score of 0.793.

The results indicate a substantial engagement bonus for content featuring performative labor, emotional manipulation tactics, and frameworks centered exclusively on financial gain. These findings suggest that audience engagement is systematically correlated with the intensive, performative efforts exerted by child creators.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC