Auditing Exposure to Harmful Content on TikTok using Multimodal Language Models: A Cross-National, Age-Stratified Study

Gemini 2.5 Flash using eight sampled frames and text achieves an aggregate kappa of 0.42 when validated against native-speaker labels for TikTok video auditing. This approach costs half as much as full video uploads, enabling a 10% sample audit of 36,971 videos for approximately $50 in API spend.