A groundbreaking study challenges the notion that AI truly understands video content. While AI models excel at identifying objects within single frames, they struggle to grasp the actions and relationships unfolding over time. Researchers have devised a clever new test, called Retrieval from Counterfactually Augmented Data (RCAD), to expose this limitation. The test involves presenting AI with a video and a set of captions. One caption accurately describes the video's action, while the others offer plausible but false descriptions. The AI's task? Pick the true caption. The results are surprising. Even state-of-the-art video AI, trained on massive datasets, performs poorly. These models are easily fooled by the alternative captions, indicating a shallow understanding of video semantics. The study highlights the difficulty of cross-frame reasoning, the ability to connect events across a sequence of frames. Humans possess this skill intuitively, but current AI struggles. To bridge this gap, researchers are exploring new approaches like “LLM-teacher,” which leverages the power of Large Language Models (LLMs) to teach AI more nuanced action recognition. This involves modifying existing video captions to create “hard negative” examples. These examples, while similar to the correct caption, contain slightly altered actions. By contrasting the video with these near-miss descriptions, the model learns to discern finer distinctions in actions. The research reveals a significant limitation in current video AI and opens a path toward models that truly 'see' and comprehend dynamic visual stories. This has profound implications for applications like video search, content moderation, and automated video analysis.