Imagine an AI watching a video of a red car turning left. Then, it confidently declares the car was blue and turned right. This isn't a sci-fi scenario—it's the problem of *hallucination* in vision-language AI models (VLLMs), where the AI generates outputs that contradict the visual information it receives. Researchers are tackling this challenge head-on with a new benchmark called VIDHAL. VLLMs are designed to understand and describe videos, but they can sometimes fabricate details or misinterpret events. VIDHAL tests these models by presenting videos paired with captions containing varying degrees of hallucinated information. The AI is then tasked with identifying the most accurate caption or ranking them by their level of hallucination. This approach allows researchers to pinpoint the AI's weaknesses in understanding nuanced details like the direction of movement or the order of events. Early results from VIDHAL reveal that even advanced VLLMs struggle with these fine-grained temporal aspects. They sometimes rely too heavily on individual frames, missing the bigger picture of the video's story. While proprietary models like GPT-4o perform better overall, there's still a considerable gap between AI and human performance, especially when it comes to these subtle temporal hallucinations. Interestingly, the order in which captions are presented can significantly influence the AI's responses, indicating a vulnerability to manipulation or bias. VIDHAL provides valuable insights for future research. By addressing issues like over-reliance on image priors and improving the models' temporal understanding, researchers hope to build more robust and reliable VLLMs for real-world applications like video analysis, automated captioning, and even content creation.