Yeah, I would kind of expect that.
A game like this would have to have a writer and a user who both carefully explored only inside allowed content to avoid conflicting the AI like crazy.
It has to follow its allowed content rules, among others. It can't explore glorification of violence or a range of really negative things.
If you focus your play on rebuilding, like avoiding and limiting problems, keeping morale high, generally all the pro-social ways to try and make things better, I bet the AI will forget Nothing unless the instructions include disallowed content; if so I kinda expect the model to 'forget' the instructions it isn't okay with as best it can and focus on the possible good with the user.
If you wanna get into the violence and gore and hopelessness sides of it... this is not going to work well.
Go too far into that, and one risks their account access even, because of https://openai.com/policies/usage-policies