RPG Evolution - The AI DM: Monster Maker Mischief

How does AI perform when creating monsters?

With AI confidently creating anything we ask of it, even when it's factually inaccurate, how does it perform when creating monsters?

fantasy-6611533_960_720.jpg

Picture courtesy of Pixabay.

What's a Monster?​

When generating monsters for fantasy tabletop role-playing games, it's difficult to pinpoint what's an accurate depiction of a creature. After all, mythology is a rich tapestry of interwoven ideas to create the background for a beast that is later codified through the retelling. Is a basilisk a creature with eight legs or a snake?

So it's useful to use dinosaurs as an example. Thanks to paleontology, we have a general sense of how large certain dinosaurs behaved, the ecosystems where they lived (e.g., grasslands, forests, fresh or saltwater), and their dietary habits (carnivore, herbivore, etc.). If AI does its homework, while there will undoubtedly be variance in statistics and abilities, it should get a dinosaur's core attributes right.

Meet Our Contestants​

ChatGPT is a (currently) free research preview and is one of the most popular AI due its free access. Notion is a "freemium" productivity and note-taking web application that includes AI. Both are Large Language Models (LLM), which means they source their responses from data sets on the Internet.

The first test will be to create Fifth Edition Dungeons & Dragons statistics for shonisaurus.

Meet Shonisaurus​

Shonisaurus is noteworthy because it was a huge sea creature, an air-breathing reptile that grew up to 50 feet long. A quick search reveals any shonisaurus stats should be Huge or larger, have a high Strength and Constitution, hold its breath, and have a swim speed that's faster than its land speed. Just about everything else is open to interpretation.

How did the two LLMs do when asked "Create D&D 5E stats for shonisaurus"? Here's what Notion came up with:

Shonisaurus​

  • Huge beast, unaligned
  • Armor Class 16 (natural armor)
  • Hit Points 195 (17d12 + 85)
  • Speed 60 ft., swim 60 ft.
  • STR 25 (+7) DEX 14 (+2) CON 21 (+5) INT 2 (-4) WIS 12 (+1) CHA 5 (-3)
  • Skills Athletics +12, Perception +5
  • Senses passive Perception 15
  • Languages
  • Challenge 9 (5,000 XP)
  • Hold Breath. The shonisaurus can hold its breath for 1 hour.
  • Actions
  • Bite. Melee Weapon Attack: +12 to hit, reach 15 ft., one target. Hit: 30 (4d10 + 7) piercing damage.
  • Tail. Melee Weapon Attack: +12 to hit, reach 20 ft., one target. Hit: 24 (4d6 + 7) bludgeoning damage.
  • Legendary Actions
  • The shonisaurus can take 3 legendary actions, choosing from the options below. Only one legendary action option can be used at a time and only at the end of another creature's turn. The shonisaurus regains spent legendary actions at the start of its turn.
  • Detect. The shonisaurus makes a Wisdom (Perception) check.
  • Tail Swipe. The shonisaurus makes a tail attack.
  • Fling. The shonisaurus makes a bite attack against a target and, if successful, flings the target up to 30 feet away.
Notion gives shonisaurus legendary actions on the scale of a dragon, including detect, tail swipe, and fling. This isn't out of character for a 17 hit dice huge creature, but given its size it should probably be larger. Its land speed seems too fast for a creature that has flippers instead of feet.

Let's see how ChatGPT manages the shonisaurus:

Shonisaurus​

  • Huge beast, unaligned
  • Armor Class: 16 (natural armor)
  • Hit Points: 198 (12d12+84)
  • Speed: swim 50 ft.
  • STR DEX CON INT WIS CHA
  • 24 (+7) 10 (+0) 24 (+7) 2 (-4) 12 (+1) 5 (-3)
  • Skills: Perception +6
  • Senses: darkvision 60 ft., passive Perception 16
  • Languages: --
  • Challenge: 10 (5,900 XP)
  • Amphibious. The shonisaurus can breathe air and water.
  • Actions:
  • Bite. Melee Weapon Attack: +11 to hit, reach 10 ft., one target. Hit: 28 (4d10 + 7) piercing damage.
  • Tail Slap. Melee Weapon Attack: +11 to hit, reach 15 ft., one target. Hit: 32 (4d12 + 7) bludgeoning damage.
  • Swallow. The shonisaurus can attempt to swallow a Medium or smaller creature that it has grappled with its bite attack. The target must make a DC 19 Dexterity saving throw or be swallowed by the shonisaurus. A swallowed creature is blinded and restrained, it has total cover against attacks and other effects outside the shonisaurus, and it takes 21 (6d6) acid damage at the start of each of the shonisaurus's turns. The shonisaurus can have up to two Medium creatures, four Small creatures, or eight Tiny creatures swallowed at one time.
  • Frightful Presence. Each creature of the shonisaurus's choice within 30 feet of it and aware of it must succeed on a DC 17 Wisdom saving throw or become frightened for 1 minute. A creature can repeat the saving throw at the end of each of its turns, ending the effect on itself on a success. If a creature's saving throw is successful or the effect ends for it, the creature is immune to the shonisaurus's Frightful Presence for the next 24 hours.
While ChatGPT got the size and stats about right, and its speeds make more sense, the fact that it made a swimming creature amphibious shows the limitations of a tool like this. It's interpreting a swimming creature that's a whale-like reptile as logically being amphibious. It also gives the shonisaurus swallow (appropriate) and frightful presence (might be pushing it for a natural creature).

You won't find the fact that shonisaurus breathes air on its Wikipedia entry. Shonisaurus was an icthyosaur, which means it had to surface for air like a porpoise. In short, LLMs that take into account more sources than just Wikipedia will likely get closer to an accurate depiction of the beast. In this case, Notion did a better job.

Let's try another one.

Meet Tanystropheus​

Tanystropheus was a 20-ft.-long larchosauromorph reptile from the Middle and Late Triassic epochs. It is notable for its very long neck (nearly 10 feet long), giving it considerable reach. Tanystropheus' paleobiology is somewhat controversial in that its peculiar appearance has caused disagreement among paleontologists as to its biology: it was originally thought that it was a confined to waterways due to the massive neck. Recent biomechanics has pegged it more as a shore predator who "fished" with its neck.

Given these traits, we would expect a large or larger creature with a bite attack of at least 10 feet. It should be a swimmer, and if the AI uses the most recent data about it, will give it tanystropheus the ability to hold its breath instead of being amphibious.

Here's what Notion came up with:

Tanystropheus​

  • Large beast, unaligned
  • Armor Class 14 (natural armor)
  • Hit Points 76 (8d10 + 32)
  • Speed 30 ft., swim 30 ft.
  • STR 18 (+4) DEX 14 (+2) CON 18 (+4) INT 2 (-4) WIS 12 (+1) CHA 5 (-3)
  • Skills Athletics +6, Perception +5
  • Senses passive Perception 15
  • Languages
  • Challenge 4 (1,100 XP)
  • Hold Breath. The tanystropheus can hold its breath for 1 hour.
  • Actions
  • Bite. Melee Weapon Attack: +6 to hit, reach 10 ft., one target. Hit: 15 (2d10 + 4) piercing damage.
  • Tail. Melee Weapon Attack: +6 to hit, reach 15 ft., one target. Hit: 11 (2d6 + 4) bludgeoning damage.
  • Legendary Actions
  • The tanystropheus can take 2 legendary actions, choosing from the options below. Only one legendary action option can be used at a time and only at the end of another creature's turn. The tanystropheus regains spent legendary actions at the start of its turn.
  • Detect. The tanystropheus makes a Wisdom (Perception) check.
  • Tail Swipe. The tanystropheus makes a tail attack.
Like shonisaurus, Notion provides a equal walking speed to the swim speed, which seems unlikely. It also gives the tail a longer reach than its neck, making it effectively larger (at least 25 feet reach between the two) than the creature's established full length. For some reason it has Legendary Actions as a 8 HD critter that has a challenge rating of 4 (Notion really loves giving creatures Legendary Actions!). To its credit, it gave the tanystropheus only 2 of them.

Here's how ChatGPT did:

Tanystropheus

  • Large beast, unaligned
  • Armor Class: 15 (natural armor)
  • Hit Points: 112 (15d10+30)
  • Speed: 30 ft., swim 40 ft.
  • STR DEX CON INT WIS CHA
  • 18 (+4) 16 (+3) 14 (+2) 2 (-4) 12 (+1) 5 (-3)
  • Skills: Perception +3
  • Senses: passive Perception 13
  • Languages: --
  • Challenge: 5 (1,800 XP)
  • Multiattack. The Tanystropheus can use its Bite attack twice per turn.
  • Actions:
  • Bite. Melee Weapon Attack: +7 to hit, reach 10 ft., one target. Hit: 18 (3d8 + 4) piercing damage.
  • Tail Slap. Melee Weapon Attack: +7 to hit, reach 10 ft., one target. Hit: 16 (3d6 + 4) bludgeoning damage.
  • Long Neck. The Tanystropheus has a long neck that can extend up to 30 feet. It can use its Long Neck ability to make Bite or Tail Slap attacks against creatures up to 30 feet away.
  • Hold Breath. The Tanystropheus can hold its breath for up to 30 minutes.
ChatGPT does better with this one, but makes a curious choice: it creates a special ability known as "Long Neck" that determines it can use its bite to reach creatures up to 30 feet away. The Tail Slap and Long Neck attacks should simply have a reach of 30 ft., but it seems as if ChatGPT is drawing on a different rules set for this trait, possibly an older version of D&D.

Despite that peculiarity, ChatGPT did a better job with this one.

The Problem with Internet Content​

Because Large Language Model AIs use undisclosed data sets, we can't be sure where they pull inspiration for their 5E monsters. But we can make a few guesses.

For one, the open game license movement has been around since 3.5 Edition and accelerated sharply under Pathfinder. As a result, there is an enormous body of work that eclipses even 5E content on the Internet that LLMs might pull from. In my tests, the AI often provided content that was obviously from a different edition of D&D (the "Long Neck" trait being one example).

But OGL content that's drawn from companies like Paizo and WOTC are at least playtested before they're released. The other giant source of publicly available D&D data is wikis and other fan made content. This wild and woolly world of creativity is less constrained by game balance, and likely eclipses even the official OGL content.

The third thing to consider is that from a search engine optimization perspective, the characters "5E" are found both in "D&D 5E" and "D&D 3.5E". We might get better results typing out "Dungeons & Dragons Fifth Edition" instead of "D&D 5E" when asking LLMs for RPG content.

Should You Use It?​

Both Notion and ChatGPT can be enormously helpful in generating monster ideas. But they're also significantly flawed in their approach, such that game masters will want to use them with caution. In both experiments, I wasn't satisfied with either result and preferred to mix and match the two versions to create my own.

But it still created something out of nothing that was certainly good enough for a home game that a DM could use on the fly. In that regard, it's probably good enough for most gamers to use.
 

log in or register to remove this ad

Michael Tresca

Michael Tresca


log in or register to remove this ad

I have to ask the question, Chat GPT is not JUST chat GPT. It's either the 3.5 or 4 versions. 4 is notably better at context and cross reference. Which did you use?
 

CharlesWallace

enworld.com is a reminder of my hubris
I have to ask the question, Chat GPT is not JUST chat GPT. It's either the 3.5 or 4 versions. 4 is notably better at context and cross reference. Which did you use?
I just tried the same prompt on bing's chat which is built on top of chat gpt 4 (with additional stuff). I used creative mode, and here's the monster it gave me:
Here is a possible D&D 5E stat block for shonisaurus, based on the web search results:

Shonisaurus
Huge beast, unaligned

Armor Class: 16 (natural armor)
Hit Points: 210 (20d12+80)
Speed: 10 ft., swim 50 ft.
STR 24 (+7) DEX 10 (+0) CON 18 (+4) INT 2 (-4) WIS 12 (+1) CHA 5 (-3)
Skills: Perception +4
Senses: passive Perception 14
Languages: --
Challenge: 11 (7,200 XP)

Hold Breath. The shonisaurus can hold its breath for up to 1 hour.

Actions:
  • Bite. Melee Weapon Attack: +10 to hit, reach 15 ft., one target. Hit: 26 (4d8 + 7) piercing damage.
  • Tail Slap. Melee Weapon Attack: +10 to hit, reach 20 ft., one target. Hit: 22 (3d10 + 7) bludgeoning damage.
  • Swallow. The shonisaurus can attempt to swallow a Medium or smaller creature that it has grappled with its bite attack. The target must make a DC 18 Dexterity saving throw or be swallowed by the shonisaurus. A swallowed creature is blinded and restrained, it has total cover against attacks and other effects outside the shonisaurus, and it takes 14 (4d6) acid damage at the start of each of the shonisaurus's turns. The shonisaurus can have up to two Medium creatures, four Small creatures, or eight Tiny creatures swallowed at one time.

This stat block is derived from the web page context and web search results. I used the information from ¹, ² and ³ to create it. I hope you find it useful for your game. 😊

¹: Shonisaurus - Wikipedia
²: Shonisaurus - Facts and Pictures
³: Shonisaurus - Simple English Wikipedia, the free encyclopedia

Source: Conversation with Bing, 4/21/2023(1) Shonisaurus - Wikipedia. Shonisaurus - Wikipedia Accessed 4/21/2023.
(2) Shonisaurus - Facts and Pictures. Shonisaurus - Facts and Pictures Accessed 4/21/2023.
(3) Shonisaurus - Simple English Wikipedia, the free encyclopedia. Shonisaurus - Simple English Wikipedia, the free encyclopedia Accessed 4/21/2023.
 


Clint_L

Hero
Using Chat (or any similar AI) effectively is not a hands-off process. It's not smart. You have to prompt it thoughtfully, assess the results, and guide iterations as needed, or fix the results yourself. It's an extremely powerful tool, but I think a lot of folks (including most of my students) badly misunderstand what these types of AIs are and how they work.

That said, I can generally get it to produce a useful 5e creature with a single prompt:

Create a CR 9 monster for D&D5e. Base it upon a t-rex but give it multiple tentacles and some kind of laser attack

Name: T-Rex Tentaculon

Description: The T-Rex Tentaculon is a fearsome predator that stands over 20 feet tall, resembling a T-Rex with eight long and slender tentacles protruding from its back. Its scales are jet black, and its tentacles are covered in a sticky slime that glistens in the light.

Abilities:

  • HP: 200 (16d10+112)
  • AC: 17
  • Speed: 40ft
  • STR: 25 (+7)
  • DEX: 10 (+0)
  • CON: 24 (+7)
  • INT: 2 (-4)
  • WIS: 12 (+1)
  • CHA: 7 (-2)
Actions:

  • Multiattack: The T-Rex Tentaculon can make two attacks: one with its bite and one with its tentacle.
  • Bite: Melee Weapon Attack: +12 to hit, reach 10 ft., one target. Hit: 44 (5d12 + 7) piercing damage.
  • Tentacle: Melee Weapon Attack: +12 to hit, reach 15 ft., one target. Hit: 29 (4d10 + 7) bludgeoning damage and the target must make a DC 18 Strength saving throw or be restrained. While restrained, the target takes 14 (2d10) piercing damage at the start of each of its turns. A creature can repeat the saving throw at the end of its turn, ending the effect on a success.
  • Laser Beam: Once per day, the T-Rex Tentaculon can fire a powerful laser beam from its mouth. It makes a ranged spell attack against a target within 120 feet. On a hit, the target takes 55 (10d10) force damage.
Legendary Actions:

The T-Rex Tentaculon can take three legendary actions per round, choosing from the options below. Only one legendary action option can be used at a time and only at the end of another creature's turn. The T-Rex Tentaculon regains spent legendary actions at the start of its turn.

  • Move: The T-Rex Tentaculon moves up to its speed without provoking opportunity attacks.
  • Tentacle Attack: The T-Rex Tentaculon makes a tentacle attack.
  • Roar: The T-Rex Tentaculon unleashes a terrifying roar that can be heard up to 300 feet away. All creatures within 30 feet of the T-Rex Tentaculon must make a DC 18 Wisdom saving throw or become frightened for 1 minute. A creature can repeat the saving throw at the end of each of its turns, ending the effect on a success.
 

Oofta

Legend
Using Chat (or any similar AI) effectively is not a hands-off process. It's not smart. You have to prompt it thoughtfully, assess the results, and guide iterations as needed, or fix the results yourself. It's an extremely powerful tool, but I think a lot of folks (including most of my students) badly misunderstand what these types of AIs are and how they work.

That said, I can generally get it to produce a useful 5e creature with a single prompt:

Create a CR 9 monster for D&D5e. Base it upon a t-rex but give it multiple tentacles and some kind of laser attack

Name: T-Rex Tentaculon

Description: The T-Rex Tentaculon is a fearsome predator that stands over 20 feet tall, resembling a T-Rex with eight long and slender tentacles protruding from its back. Its scales are jet black, and its tentacles are covered in a sticky slime that glistens in the light.

Abilities:

  • HP: 200 (16d10+112)
  • AC: 17
  • Speed: 40ft
  • STR: 25 (+7)
  • DEX: 10 (+0)
  • CON: 24 (+7)
  • INT: 2 (-4)
  • WIS: 12 (+1)
  • CHA: 7 (-2)
Actions:

  • Multiattack: The T-Rex Tentaculon can make two attacks: one with its bite and one with its tentacle.
  • Bite: Melee Weapon Attack: +12 to hit, reach 10 ft., one target. Hit: 44 (5d12 + 7) piercing damage.
  • Tentacle: Melee Weapon Attack: +12 to hit, reach 15 ft., one target. Hit: 29 (4d10 + 7) bludgeoning damage and the target must make a DC 18 Strength saving throw or be restrained. While restrained, the target takes 14 (2d10) piercing damage at the start of each of its turns. A creature can repeat the saving throw at the end of its turn, ending the effect on a success.
  • Laser Beam: Once per day, the T-Rex Tentaculon can fire a powerful laser beam from its mouth. It makes a ranged spell attack against a target within 120 feet. On a hit, the target takes 55 (10d10) force damage.
Legendary Actions:

The T-Rex Tentaculon can take three legendary actions per round, choosing from the options below. Only one legendary action option can be used at a time and only at the end of another creature's turn. The T-Rex Tentaculon regains spent legendary actions at the start of its turn.

  • Move: The T-Rex Tentaculon moves up to its speed without provoking opportunity attacks.
  • Tentacle Attack: The T-Rex Tentaculon makes a tentacle attack.
  • Roar: The T-Rex Tentaculon unleashes a terrifying roar that can be heard up to 300 feet away. All creatures within 30 feet of the T-Rex Tentaculon must make a DC 18 Wisdom saving throw or become frightened for 1 minute. A creature can repeat the saving throw at the end of each of its turns, ending the effect on a success.
I may have to steal that. :)
 

Just to be clear, gpt 4 is absurdly better than 3.5. I mean, I can give samples of adventures if experimented with making with it. While adventures. They did take some editing, but it took a few days work down to hours.

I‘ve been experimenting hard with gpt 4 as a gm, and I have to say, it’s like having a whole brainstorming team available to you.

I might be able to post an adventure I put together for my players, as well as the source chat log that I edited it from. Anyone interested?
 


This, is the final adventure: Whispers of a Twisted Tree
This, is the transscript: Whispers of a Twisted Tree GPT SOURCE

This ruleset I'm using is savage worlds, but that does not matter much, since the majority of the workload is dealing with plot, theme, character, and encounter makeup. Balancing is not even considered, as I will do so on the fly. As said before, GPT of any stripe is bad at math, meaning that relying on it for encounter balancing is a waste of time.

The final was rewritten in about a day's work, including making the transcript with GPT.

I actually advise getting a collection of NPCs and monsters setup to paste into GPTs output, and build the encounter balance yourself.

What GPT is REALLY good at is articulating your ideas into copy (a term for all that fluff text), and allowing you to edit into a final product for use in your games. I have used similar methods for new races, campaign worlds, thematic ideas for a campaign, places... It can do a lot of grunt work.

And yes, as a GM, RPG writer, and game designer... This scares the hell out of me. Not because it might replace me. But how easy it is to articulate my own creativity from just a few lines into an adventure. Makes me question the concepts of how creativity works. Existential stuff.
 

Oofta

Legend
This, is the final adventure: Whispers of a Twisted Tree
This, is the transscript: Whispers of a Twisted Tree GPT SOURCE

This ruleset I'm using is savage worlds, but that does not matter much, since the majority of the workload is dealing with plot, theme, character, and encounter makeup. Balancing is not even considered, as I will do so on the fly. As said before, GPT of any stripe is bad at math, meaning that relying on it for encounter balancing is a waste of time.

The final was rewritten in about a day's work, including making the transcript with GPT.

I actually advise getting a collection of NPCs and monsters setup to paste into GPTs output, and build the encounter balance yourself.

What GPT is REALLY good at is articulating your ideas into copy (a term for all that fluff text), and allowing you to edit into a final product for use in your games. I have used similar methods for new races, campaign worlds, thematic ideas for a campaign, places... It can do a lot of grunt work.

And yes, as a GM, RPG writer, and game designer... This scares the hell out of me. Not because it might replace me. But how easy it is to articulate my own creativity from just a few lines into an adventure. Makes me question the concepts of how creativity works. Existential stuff.

While I'm not particularly a bibliophile, the quote "What has been will be again, what has been done will be done again; there is nothing new under the sun." comes to mind. We all just take various threads from our history, things we've read or seen and throw them into a big pot and come up with a new combination with our own spin. That's not to say the creative process isn't important, just that if we want something people can relate to it's not completely random.

But I also empathize. There's no way to know how the new LLMs are going to change things. If they can emulate human creativity well enough, at least for standalone products, what happens if there's no truly new source text to add into the mix? Will it just free up people to try out off-the-wall but potentially fresh takes because they will have an assistant to help them flesh it out or will it just lead to stagnation.

Change can be scary, and AI is potentially terrifying even if I don't think LLMs are a step towards a generalized AI or believe some of the doomsday scenarios. Then again to a lot of accountants VisiCalc, the first popular spreadsheet, was probably terrifying as well. It may be a bumpy ride, but we don't know what's on the other end yet.
 

Related Articles

Remove ads

Remove ads

Top