Well, no. Models are not “self-learning”. This is unrelated to size of the corpus. In fact, it was proven times and times again, that poisoning a model is not a costly proces.