It isn't actually, this is a separate layer of censorship ontop of the actual deepseek model.
If you run the model locally via ollama it won't output answers like that, it'll basically just act like a chinese official broadcast live on BBC who has been firmly instructed to avoid outright lies.