Classifying TikToks Locally: Political Content Detection with Phi-4 on Android

Large Language Models (LLMs) like ChatGPT-4o and Phi-4 have demonstrated great potential in identifying complex semantic content, such as political discourse, especially in cloud or desktop-based settings. However, their application on mobile devices, where privacy concerns are high, remains largely unexplored. Mobile phones have become the main way people access political discussions and news. This study investigates whether local execution of LLMs can serve as a viable, privacy-preserving way for analyzing political content seen on mobile screens. Using a Google Pixel 9, we benchmarked Phi-4 with 2,000 OCR-extracted text samples from TikTok screen recordings, comparing classification latency and feasibility. While results show that local classification is possible, latency is high, averaging over 14 seconds per sample. Although dictionary-based methods are faster, they lack the semantic flexibility of LLMs. Our findings suggest a hybrid approach and targeted frame selection strategies could enable scalable, privacy-friendly mobile media analysis in the near future.