Someone has called this 'semantic bleeding' which I think is an amazing name. It also happens in text generated by large language models.