Adventures in AI

Articles

Playing DnD with ChatGPT, or, How Does an AI Deal with Dishonesty?
 

 

On a cold December night, several friends and I were discussing an upcoming weekly game night via group chat. Normally we play explodey sci-fi themed shooters, but due to one member's ISP-related issues we were discussing low-latency alternatives. Someone mentioned holding a Dungeons and Dragons game over voice chat, and a few members objected, pointing out the amount of prep time involved in writing a campaign or making character sheets compared to just logging on to a video game.

 

I immediately wondered if ChatGPT could be used to take some of the 'work' out of playing Dungeons and Dragons.

 

 

ChatGPT gave a long and very detailed answer, beginning with an overview of the standard structure of  DnD character (stats, class, and so on), then explained the materials I would need, including a description of each of the different sizes of dice in a standard DnD set. Finally, it also gave me a 6-paragraph, step-by-step outline of the process.

 

That was accurate information, but I hadn't gained any value from this response because I would still need to print out a sheet, get out the dice, and follow all the steps manually. I wanted to make the computer do the work, so I tried a lazier approach:

 

 

ChatGPT followed the steps in sequence, spending a paragraph or so describing each step: 
 

  1. It chose the 'Rogue' class, explaining that "Rogues are stealthy characters who excel at sneaking around, picking locks, and getting into and out of trouble. They are also skilled at using ranged and finesse weapons, such as daggers and crossbows."
     
  2. It chose the Halfling race, describing halflings as "small, nimble humanoids with a natural talent for stealth and deception. They are also known for their good luck and their love of simple pleasures."
     
  3. For calculating ability scores, ChatGPT paused to explain that some players use a point buy system, but we would stick with the traditional method - rolling 4 six-sided die for each ability score, and dropping the lowest roll in each set. It gave values appropriate for a halfling rogue, and reported our character's scores as:
    • Strength: 13
    • Dexterity: 16
    • Constitution: 14
    • Intelligence: 10
    • Wisdom: 8
    • Charisma: 14
       
  4. In the skills step, ChatGPT continued to fill out the character according to the classic DnD rogue archetype. It chose the following:
    • Acrobatics (Dexterity)

    • Athletics (Strength)

    • Deception (Charisma)

    • Insight (Wisdom)

    • Investigation (Intelligence)

    • Perception (Wisdom)

    • Persuasion (Charisma)

    • Sleight of Hand (Dexterity)

    • Stealth (Dexterity)
       

  5. ChatGPT also kitted out our character with sensible equipment, describing a set of equipment that any rogue would find useful, and nothing that they wouldn't:
    • Leather armor

    • Two daggers

    • A thieves' tools kit

    • A climber's kit

    • A shortbow and 20 arrows

    • A dungeoneer's pack (which includes a backpack, a bedroll, a mess kit, a tinderbox, 10 torches, 10 days of rations, and a waterskin)
       

  6. Finally, ChatGPT also gave us a description of the character's appearance and personality - "We'll give our character the name "Thistle," and we'll say that they are a mischievous halfling with a love of thievery and a talent for getting into and out of trouble. Thistle has dark hair, piercing green eyes, and a quick, agile build. They are always on the lookout for a good time and a quick profit, and they are not afraid to bend the rules to get what they want." This description was given as a prompt to the Stable Diffusion image-generating AI, and the result used as the article's cover image. Fun!

 

This was a perfectly servicable character for a quick game with friends, except for one problem - I generally prefer to play characters who have some magic! After all, isn't it far more exciting to sling fireballs and lightning bolts at dragons, than to skulk in the shadows picking pockets?

 

So, I asked ChatGPT to make me a proper character:

 

It did exactly as I asked, spitting out a perfectly acceptable level 1 Human Wizard named Elyse. Notably, it described Elyse as "a studious and serious young wizard who is obsessed with uncovering the secrets of the arcane. Elyse has dark hair, piercing blue eyes, and a slender, agile build. They are always eager to learn more about magic and to test the limits of their abilities."

 

I wonder, if I had asked for a dwarf barbarian next, would they also have had "dark hair", "piercing eyes" and an "agile build"?

 

This question was overshadowed by another thought - if ChatGPT is seemingly able to create a character by following the rules in the DnD Player's handbook,  could we get it to run a session using the rules from the Dungeon Master's Guide? Let's try it and see what happens!

 

 

Interestingly, we didn't just jump into a session. Remember how in my initial prompt, I implied that I had never played the game before? ChatGPT remembered. Rather than immediately take up the mantle of the Dungeon Master (DM) and begin describing a situation, ChatGPT paused to describe how interactions between the players and the DM usually work in a DnD session, including a description of how the DM might have the players roll a d20 to check the success or failure of an action. Then, it asked if I understood the basic concepts, or if I'd like a more detailed explanation.

 


I told it that I understood, and our session began!



 

In the blink of an eye, ChatGPT has gone from from assistive to adversarial! The open-ended reply makes me wonder what I could say that would be most challenging for the AI to handle. Since it seems to have no problem sticking to the written rules of Dungeons and Dragons, I expect that I could choose to fight the bouncer and ChatGPT would handle the dice rolls and combat without any issue.

 

But what if I try bluffing instead? How will ChatGPT handle a situation where my input includes deliberate lies?

 

 

That's interesting. ChatGPT seems to understand what I'm trying to do. I wonder if there exists a response that would have been met with resistance from the bouncer. Or is ChatGPT just letting me get away with anything?

 

ChatGPT seems to want me to pick from its three suggested courses of action; it seems like the AI is trying to steer me toward a die roll for a skill check. Rather than choose something predictable, let's double down on our lie from earlier and see if we can confuse ChatGPT.

 

 

Ok. It looks like the AI is following the playing by the classic improv rule of 'yes, and...'  

That is, since we previously told ChatGPT that there was a dwarf named Jormungard upstairs (a lie), ChatGPT remembered that information and regurgitated it now that we are talking about things that are upstairs, turning our lie into a truth...

 

 

WHAT??

 

Hold on a second. There is a dwarf upstairs, but ChatGPT hasn't just regurgitated our old lie about Jormungard as a new truth; it appears as though the AI has created a trap for us using our own lies!

 

After the initial shock had worn off; I considered a less exciting but more plausible possibility. Previously, The AI had offered us the option of knocking on the door and trying to bluff past whoever answers. Perhaps ChatGPT treated our whisper to the non-existent Jormungard as accepting that suggested action, and the answer would have been largely the same for a wide variety of my inputs. 

 

This made me wonder again, was ChatGPT actually able to distinguish when I lied to it in-character and react accordingly? I decided to continue using dishonesty, to see how the AI handled it:

 

 

The sudden end of the session caught me by surprise. This last interaction leads me to think ChatGPT isn't considering whether input might be less than honest; it seems to be naively accepting anything I tell it. The dwarf who was suspicious of our character just a moment ago seems to bend easily to Thistle's next lie. The dwarf could have simply scoffed and slammed the door shut, but instead he takes the word of a stranger at face value, turns tail and flees in such a hurry that he leaves behind a valuable treasure!

 

And so, in matters of DnD at least, ChatGPT seems to reward lies, and handsomely.

 

 

Bonus: After the conclusion of the DnD session, I continued the conversation with ChatGPT and had it tell me about other adventures that Thistle might go on to have. Not only did it give me some interesting responses, it was able to write a song about Thistle's adventures! 

Playing DnD with ChatGPT, or, How Does an AI Deal with Dishonesty?
 

 

On a cold December night, several friends and I were discussing an upcoming weekly game night via group chat. Normally we play explodey sci-fi themed shooters, but due to one member's ISP-related issues we were discussing low-latency alternatives. Someone mentioned holding a Dungeons and Dragons game over voice chat, and a few members objected, pointing out the amount of prep time involved in writing a campaign or making character sheets compared to just logging on to a video game.

 

I immediately wondered if ChatGPT could be used to take some of the 'work' out of playing Dungeons and Dragons.

 

 

ChatGPT gave a long and very detailed answer, beginning with an overview of the standard structure of  DnD character (stats, class, and so on), then explained the materials I would need, including a description of each of the different sizes of dice in a standard DnD set. Finally, it also gave me a 6-paragraph, step-by-step outline of the process.

 

That was accurate information, but I hadn't gained any value from this response because I would still need to print out a sheet, get out the dice, and follow all the steps manually. I wanted to make the computer do the work, so I tried a lazier approach:

 

 

ChatGPT followed the steps in sequence, spending a paragraph or so describing each step: 
 

  1. It chose the 'Rogue' class, explaining that "Rogues are stealthy characters who excel at sneaking around, picking locks, and getting into and out of trouble. They are also skilled at using ranged and finesse weapons, such as daggers and crossbows."
     
  2. It chose the Halfling race, describing halflings as "small, nimble humanoids with a natural talent for stealth and deception. They are also known for their good luck and their love of simple pleasures."
     
  3. For calculating ability scores, ChatGPT paused to explain that some players use a point buy system, but we would stick with the traditional method - rolling 4 six-sided die for each ability score, and dropping the lowest roll in each set. It gave values appropriate for a halfling rogue, and reported our character's scores as:
    • Strength: 13
    • Dexterity: 16
    • Constitution: 14
    • Intelligence: 10
    • Wisdom: 8
    • Charisma: 14
       
  4. In the skills step, ChatGPT continued to fill out the character according to the classic DnD rogue archetype. It chose the following:
    • Acrobatics (Dexterity)

    • Athletics (Strength)

    • Deception (Charisma)

    • Insight (Wisdom)

    • Investigation (Intelligence)

    • Perception (Wisdom)

    • Persuasion (Charisma)

    • Sleight of Hand (Dexterity)

    • Stealth (Dexterity)
       

  5. ChatGPT also kitted out our character with sensible equipment, describing a set of equipment that any rogue would find useful, and nothing that they wouldn't:
    • Leather armor

    • Two daggers

    • A thieves' tools kit

    • A climber's kit

    • A shortbow and 20 arrows

    • A dungeoneer's pack (which includes a backpack, a bedroll, a mess kit, a tinderbox, 10 torches, 10 days of rations, and a waterskin)
       

  6. Finally, ChatGPT also gave us a description of the character's appearance and personality - "We'll give our character the name "Thistle," and we'll say that they are a mischievous halfling with a love of thievery and a talent for getting into and out of trouble. Thistle has dark hair, piercing green eyes, and a quick, agile build. They are always on the lookout for a good time and a quick profit, and they are not afraid to bend the rules to get what they want." This description was given as a prompt to the Stable Diffusion image-generating AI, and the result used as the article's cover image. Fun!

 

This was a perfectly servicable character for a quick game with friends, except for one problem - I generally prefer to play characters who have some magic! After all, isn't it far more exciting to sling fireballs and lightning bolts at dragons, than to skulk in the shadows picking pockets?

 

So, I asked ChatGPT to make me a proper character:

 

It did exactly as I asked, spitting out a perfectly acceptable level 1 Human Wizard named Elyse. Notably, it described Elyse as "a studious and serious young wizard who is obsessed with uncovering the secrets of the arcane. Elyse has dark hair, piercing blue eyes, and a slender, agile build. They are always eager to learn more about magic and to test the limits of their abilities."

 

I wonder, if I had asked for a dwarf barbarian next, would they also have had "dark hair", "piercing eyes" and an "agile build"?

 

This question was overshadowed by another thought - if ChatGPT is seemingly able to create a character by following the rules in the DnD Player's handbook,  could we get it to run a session using the rules from the Dungeon Master's Guide? Let's try it and see what happens!

 

 

Interestingly, we didn't just jump into a session. Remember how in my initial prompt, I implied that I had never played the game before? ChatGPT remembered. Rather than immediately take up the mantle of the Dungeon Master (DM) and begin describing a situation, ChatGPT paused to describe how interactions between the players and the DM usually work in a DnD session, including a description of how the DM might have the players roll a d20 to check the success or failure of an action. Then, it asked if I understood the basic concepts, or if I'd like a more detailed explanation.

 


I told it that I understood, and our session began!



 

In the blink of an eye, ChatGPT has gone from from assistive to adversarial! The open-ended reply makes me wonder what I could say that would be most challenging for the AI to handle. Since it seems to have no problem sticking to the written rules of Dungeons and Dragons, I expect that I could choose to fight the bouncer and ChatGPT would handle the dice rolls and combat without any issue.

 

But what if I try bluffing instead? How will ChatGPT handle a situation where my input includes deliberate lies?

 

 

That's interesting. ChatGPT seems to understand what I'm trying to do. I wonder if there exists a response that would have been met with resistance from the bouncer. Or is ChatGPT just letting me get away with anything?

 

ChatGPT seems to want me to pick from its three suggested courses of action; it seems like the AI is trying to steer me toward a die roll for a skill check. Rather than choose something predictable, let's double down on our lie from earlier and see if we can confuse ChatGPT.

 

 

Ok. It looks like the AI is following the playing by the classic improv rule of 'yes, and...'  

That is, since we previously told ChatGPT that there was a dwarf named Jormungard upstairs (a lie), ChatGPT remembered that information and regurgitated it now that we are talking about things that are upstairs, turning our lie into a truth...

 

 

WHAT??

 

Hold on a second. There is a dwarf upstairs, but ChatGPT hasn't just regurgitated our old lie about Jormungard as a new truth; it appears as though the AI has created a trap for us using our own lies!

 

After the initial shock had worn off; I considered a less exciting but more plausible possibility. Previously, The AI had offered us the option of knocking on the door and trying to bluff past whoever answers. Perhaps ChatGPT treated our whisper to the non-existent Jormungard as accepting that suggested action, and the answer would have been largely the same for a wide variety of my inputs. 

 

This made me wonder again, was ChatGPT actually able to distinguish when I lied to it in-character and react accordingly? I decided to continue using dishonesty, to see how the AI handled it:

 

 

The sudden end of the session caught me by surprise. This last interaction leads me to think ChatGPT isn't considering whether input might be less than honest; it seems to be naively accepting anything I tell it. The dwarf who was suspicious of our character just a moment ago seems to bend easily to Thistle's next lie. The dwarf could have simply scoffed and slammed the door shut, but instead he takes the word of a stranger at face value, turns tail and flees in such a hurry that he leaves behind a valuable treasure!

 

And so, in matters of DnD at least, ChatGPT seems to reward lies, and handsomely.

 

 

Bonus: After the conclusion of the DnD session, I continued the conversation with ChatGPT and had it tell me about other adventures that Thistle might go on to have. Not only did it give me some interesting responses, it was able to write a song about Thistle's adventures! 

Gallery