A Scalable Sampling Approach for Artificial Intelligence‐Based Alcohol Content Estimation in Movies

INTRODUCTION: The growing accessibility of movies through streaming platforms has expanded audience reach, but also increases exposure to alcohol portrayals, which is an established risk factor for alcohol use. Hence, estimating alcohol depictions is important yet challenging due to the time and labor involved. Artificial Intelligence offers a scalable solution for analysing movie frames; however, processing every frame of a full-length movie at 25 frames per second (fps) requires extensive computational resources. Thus, we aimed to test whether lower-frequency sampling would affect the accuracy of alcohol exposure estimates. METHODS: We analysed 20 feature-length movies with varying known alcohol visibility and analysed each frame using zero-shot predictions from a LLaVA v1.6 model (accuracy = 95%) as our baseline. We applied uniform downsampling from 25 fps (full-framerate) to 1 fps and sparse interval sampling of 1 frame per N seconds (N = 1,2,…,10), measuring both alcohol-proportion estimates and execution time. To assess the sampling-induced error, we computed the difference score, as the difference between sampled and full-frame alcohol proportions. RESULTS: A sampling frequency of 1 fps yielded an average difference score below 0.10 compared to the full-frame analysis, while reducing execution time by 25-fold. Error increased at sparser intervals, reaching a difference score of 0.46 at one frame per 10 s. DISCUSSION AND CONCLUSION: Reducing the sampling frequency from 25 to 1 fps resulted in only a minimal loss of accuracy but a substantial reduction in execution time. This finding supports 1 fps as a practical and scalable sampling frequency for large-scale movie alcohol exposure estimation.