back to article Hallucinating ChatGPT finds a role playing Dungeons & Dragons

Boffins have found a role for AI chatbots where habitual hallucination isn't necessarily a liability. The eggheads – based at at the University of Pennsylvania and the University of Maryland, Baltimore County, in the US – enlisted OpenAI's large language models (LLMs) to help with fantasy role playing, specifically Dungeons & …

  1. Bronk's Funeral

    FlowGPT have been running game jams along these lines for a while now. I've been trying to put together a prompt that makes GPT act as a D&D *player* that a user can act as DM with—it's hard to wrangle but the sodding thing goes off in some wild directions, sometimes. It's fun, because it's almost entirely unpredictable, and a good DM should be resilient to unpredictability.

    Here's an early attempt. Ignore the first wall of text—it's the prompt: https://chat.openai.com/share/3ac574f6-c49f-4b80-9e41-317487f69f2f

    1. ExampleOne

      it's hard to wrangle but the sodding thing goes off in some wild directions, sometimes.

      Only sometimes? So it is marginally more predictable than a human player?

      1. ThatOne Silver badge
        Happy

        > more predictable than a human player?

        Depends on the player. My (totally human) players were highly predictable: Find a good tavern, eat & drink and chase the stable boy (note players were female, playing female characters). :-D

        I always had the direst problems getting them out of that tavern. "What? Mysterious things happen nearby? Yes, we probably should look into it - next week." "What? Fabulous treasure? Yes, sounds enticing, but now it's dinner time." I finally solved the problem by making a never-ending "campaign chain" where there was always still something absolutely essential to do: No time for taverns, duty calls...

        1. big_D Silver badge

          We had a sex-obsessed character in one game. The DM was getting tired of him wenching his way through every tavern or encounter on the road, so he had made one encounter with witch, who cursed him...

          Let's just say, if it was dark, we didn't need a torch, just a trouser-less dwarf. It also made the character less appealing to the women he encountered...

    2. Anonymous Coward
      Anonymous Coward

      Ah, so it plays chaotic neutral characters?

  2. b0llchit Silver badge
    Childcatcher

    Quantized magic

    Not hallucination, it is magic. And soon, the interface will allow bolts of magic to exit the boxes and screens on a roll of the decahedraspindletron approaching the value of afternoon thee with a penetration factor of deadbeforeyouknowit and an uncertainty value of nininsh.

  3. ExampleOne

    Interestingly this was my first thought for a good application of chatgpt, with a human DM keeping factual statements correct.

    The hallucinations aren’t a problem in NPC conversations, because the NPC could just be wrong!

    1. Version 1.0 Silver badge
      Childcatcher

      ..."a good application of chatgpt" these days is like the mid 60's when many people said that bananas or watermelons was cheaper and easier than dating with the option of pleasure or just a bite to eat.

      As a kid we all learned how to play chess which showed everyone how to predict what the other player was going to try and do - that was a subtle but very effective lesson teaching us all the variations in peoples reactions to our actions. I see all apps like ChatGPT as just getting things done, not doing anything much that creates a deep learning environment for kids.

  4. Arthur the cat Silver badge

    CALYPSO

    Collaborative Assistant for Lore and Yielding Plot Synthesis Objectives

    Which leads to the question: can LLMs produce better contrived acronyms than humans?

    1. Anonymous Coward
      Anonymous Coward

      Re: CALYPSO

      Spurred by your comment, I decided to put this to the test. I’m developing a new system for work, and I have a name for it that I think is pretty clever, but I had a hard time coming up with a suitable, not-too-cheesy backronym for it. I registered an account with OpenAI and gave ChatGPT a pretty short prompt, and it came back with a reasonable suggestion right off the bat. It did however want to spell “Control” with a K, so I asked it to stop that nonsense. It came back with another decent attempt that had made some incorrect (but not all that improbable) assumptions; I pointed those out and the third attempt was just… perfect.

      So… maybe?

  5. t245t
    Boffin

    D&D first appeared in 1974 as a role-playing game ..

    D&D first appeared in 1974 as a role-playing game (RPG) in which players assumed the roles of adventuring medieval heroes and acted out those personalities under a storyline

    It was the only reason I signed up to college /s

  6. EvilDrSmith Silver badge

    Wrong game

    Surely they shouldn't have used D&D, they should have been running 'Paranoia'.

    Trust the Computer, the Computer is your friend....

    1. Julz

      Re: Wrong game

      Your never alone with a clone...

  7. Michael Wojcik Silver badge

    God forbid we be creative

    Yes, let's automate the fun, imaginative parts of the game, so we can concentrate on rolling those dice.

    On the other hand, LLMs, with their access to huge corpora of text and data, may be able to help our research DMs study the fundamental questions of D&D. Maybe we'll finally be able to prove that PC ≠ NPC.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like