Researchers from France, Italy, and Sweden conducted an audit of TikTok's content using multimodal language models to identify exposure to harmful content among young users. They created sockpuppet accounts representing different age groups and collected over 36,000 videos through passive scrolling and active sessions. The team validated four language models against native-speaker labels and found that one model, Gemini 2.5 Flash, performed best in detecting harmful content at a lower cost than traditional annotation methods. The study revealed varying rates of exposure to harm across different countries and age groups, with Italy showing the highest rate of harm among all age groups. --- Why it matters: This research matters because it provides a scalable approach for cross-national youth-safety audits on social media platforms like TikTok, which can help identify areas where young users are most at risk. The findings can inform policymakers and platform moderators about effective strategies to reduce exposure to harmful content. Source: https://arxiv.org/abs/2608.17583