A key limitation of current video models is generating realistic counterfactual scenarios, such as failures