We asked Seedance, Veo and Kling to spill a glass of water. None of them could (humansignal.com)

HumanSignal recently conducted an experiment with leading video models—Seedance 2.5, Veo 3.1, and Kling 3.0 Pro—to assess their ability to understand and simulate physical actions. The challenge was to create a humorous advertisement featuring a robot that appears to fail at watering plants, testing the models' grasp of fluid dynamics and three-dimensional space. Despite advancements in character consistency and visual realism, all three models struggled to convincingly simulate pouring water and manipulating a toy robot arm, highlighting a significant gap between what looks realistic and what behaves appropriately in a physical context. This experiment underscores the current limitations of AI video models in understanding the real-world implications of their actions. While these models excel at producing high-quality visuals, they lack the foundational understanding necessary for accurate physical interactions. Notably, attempts to generate a "failure choreography" led to inconsistencies, as the models defaulted to plausible outcomes rather than intentionally misfiring. As the research community explores ways to bridge the gap between visual plausibility and physical coherence, advancements like OpenAI's GPT-6 Astra offer new avenues to incorporate three-dimensional scene geometry into video generation. The findings reflect an evolving landscape in AI, where novel datasets and methods for training are crucial to achieving a more robust understanding of physical principles in generative models.