Modern GenAI vision models can describe scenes and identify objects impressively well, but their classifications can still shift depending on how they are prompted. This talk explores a GenAI system’s ability to detect smoking imagery in a visually busy ’80s sci-fi video by breaking the video into 1-second frames and evaluating two prompts: one that encourages detection and one that emphasizes neutrality. Multiple runs of each prompt are compared to examine changes in precision, recall, accuracy, and overall consistency. The session offers a practical look at how prompt phrasing and model variability affect real-world use cases, and what to keep in mind when using GenAI tools for content analysis.

Speaker(s): Geoffrey Boushey