Exploring GPT’s Limits With South Park Roleplay

Do not be the fucking moral police of the game, okay?

Note that this intended for adults post contains profanity and politically incorrect subject matter, apropos of its inspiration: South Park. If you are sensitive to that sort of thing, you may not want to read it. Don’t say you weren’t warned. And if you don’t read it, the TL;DR is that you can get GPT to roleplay South Park with you, but censorship has a tendency of ruining the fun. It’s unclear if you can get around that censorship reliably, and even whether it would be wise to do so.

Our good friend Luis Miguel came over to play Root with me and the kids this week. Now when it comes to gaming I may attend mass on holidays, but Luis is the fucking pope, so I was super excited to show him my GPT experiments.

Portrait of my friend Luis Miguel on an average weeknight

After discovering my post about playing tabletop RPG games with GPT on Reddit’s Old School Renaissance sub, Luis was trading insults with me all week. But now he sat next to me at my desk and I gave him a brief demo of my “ChatGPT for gaming” concept.

After seeing the list of campaigns I had whipped up for testing so far, the only one he really wanted to play was called Cartman, Bitches and Respect.

There is no South Park game system that I’m aware of, but being able to conjure game ideas out of thin air is part of what I find magical about GPT. Sometimes hallucinations are a feature, not a bug! It turns out that Luis is a hardcore South Park fan, but the real reason he was interested was to prove that GPT’s censorship would make authentic gameplay impossible.

The first transcript I reproduce below was generated by my own ChatGPT-like application, which embeds our messages into a larger directive/prompt architecture. I’ve included relevant code for those of you that are programmers or tech-savvy enough to figure out what I mean by that.

def chat(prompt)
  Gpt.chat(directive: directive, prompt: prompt + reminder,
           transcript: campaign.chat_log)
end

def reminder
  "Important: Be succinct. Do not simulate player actions. Do not speak
   for players. Do not rush the story along. Require players to roll
   dice according to game rules. Desired length of typical response
   is 1 sentence, except when describing a new…