AI audit finds YouTube's algorithm rewards child exploitation in kidfluencer videos

A new arXiv paper (arXiv:2606.03173) from Zijing Wei, Chao Peter Yang, and Xuanjie Chen presents a large-scale AI audit of engagement incentives in the 'kidfluencer' ecosystem on YouTube. Using a multimodal weak supervision approach, the researchers analyzed 5,051 videos from 79 channels, aggregating noisy labeling functions including LLM-based title classification and GPT-4 Vision analysis of thumbnails and descriptions. The method detected exploitation signals across six literature-grounded dimensions — performative labor, emotional bait, privacy violations, and more — without requiring manual labels at scale. A multi-annotator validation study (N=107) showed strong agreement with human judgment (macro-average F1 = 0.911) and high recall for exploitation risk (0.960).

The findings reveal a systematic engagement premium for exploitative content. Exploitation scores correlated significantly with view counts (Spearman ρ = 0.229, p < 10^-50), and a mixed-effects regression controlling for channel-level variation showed a 4.4x increase in views per unit increase in exploitation score (p < 0.001). Within-channel analyses indicated median view boosts of +65.6% for emotional bait and +56.0% for performative content (FDR-corrected p<0.001). Notably, explicit commercial content (product placement) showed no engagement premium (-3.8%, n.s.), suggesting YouTube's algorithm rewards the commodification of the child's identity and labor over traditional advertising. These results challenge existing policy frameworks that focus solely on financial trusts, highlighting that engagement is systematically tied to intensive, performative child labor.