Generative tooling makes camera movement free, and free movement gets applied to everything, which is why so much generative video has the same lazy drift on every shot.
A specified path is a decision: this shot pushes in because the story pushes in, that one is locked off because a locked frame among moving ones reads as confidence. Naming the move in camera-department language rather than as "cinematic" is most of the work.
A path also compounds with subject motion, and the compound is what breaks coherence. A moving camera on a moving subject is the most expensive thing you can ask for and the first place to economise when a clip will not hold.
Why does all my generated video have the same slow push-in?
Because camera movement is free and gets applied by default. Specify the move per shot, and specify locked-off where the shot does not need one.
Why does a moving camera make coherence worse?
It compounds with subject motion, so the model has to keep two changing relationships consistent at once. Locking the camera is the cheapest way to buy clip length.