A new study published on arXiv details a method for auditing harmful content exposure on TikTok using multimodal large language models (MLLMs). Researchers found that Gemini 2.5 Flash, when analyzing video frames and text, performed best among tested LLMs for this task. The study utilized sockpuppet accounts across different age groups in France, Italy, and Sweden, revealing that keyword searches significantly increased the detection of harmful content compared to passive scrolling, though this effect was temporary. Italy showed the highest rates of harmful content exposure across all ages under passive scrolling conditions. AI
IMPACT Provides a scalable method for auditing online platform safety and content exposure across different demographics and regions.
RANK_REASON Academic paper detailing a new methodology for content auditing. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
