Well, with models like SD at least, the datasets are large enough and the employees are few enough that it is impossible to have a human filter every image. They scrape them from the web and try to filter with AI, but there is still a chance of bad images getting through. This is why most companies install filters after the model as well as in the training process.