Negative prompts do not fix motion glitches in AI-generated video because the guidance that suppresses unwanted content in a single frame does not suppress it consistently across a sequence of frames. Negative prompting works by steering the model away from a second, unwanted version of the same image, not by policing motion over time. The technique rides on classifier-free guidance (CFG), introduced by Ho and Salimans in 2022 and popularized in Stable Diffusion 2.0, where the community figured out you could extrapolate away from whatever the negative text described. Video breaks that assumption because a clip is dozens of frames generated under a shared temporal model. New research on physical plausibility in video diffusion identifies two failure modes: a lagged suppression effect, where unwanted behavior gets pushed down in some frames but creeps back in later ones, and cumulative trajectory bias, where small deviations compound frame over frame until the fan, the mouth, or the cable is doing something physically impossible by the end of the clip. Researchers behind a 2025 paper built a workaround called Synchronized Decoupled Guidance because standard negative prompting couldn’t suppress implausible motion “immediately and consistently throughout denoising.” A negative prompt tells the model what a frame should not look like, but says nothing about what a sequence of frames should not do. Over-constrained negative prompts can amplify the very flaws they’re meant to eliminate.