Deep Dungeons and Deeper Neural Networks: Using ChatGPT in Your TTRPG

Old photo of a Babbage Difference Engine.

One of the main reasons I wanted to do an RPG blog was to discuss my experience with using ChatGPT as a DM.

About six months ago I was informed by my good friend, an IT specialist, that my career as a business writer was basically over. I was going to be replaced with AI, and my only hope of saving my career was to understand AI myself.

In a panic, I learned to use, and experimented with, a number of LLMs and image generators, mostly but not exclusively ChatGPT and Midjourney. And I learned a lot. The main thing I learned was that my job was in no immediate danger.

Not Exactly Zora from Star Trek: Discovery: What Is an LLM?

ChatGPT is a large language model (LLM). To oversimplify, a computer program examined the entire internet and as many digital texts as the programmers could acquire, and built a multidimensional map of key words from these documents, based on the words’ relationships with other words. This creates a lossy (incomplete) database of all this information, storing a kind of synopsis of a massive amount of data as a much smaller “model.” Using this model, algorithms can appear to understand questions (“prompts”), and produce comprehensible answers, based only on the relationships between words.

It is correct, technically, to refer to this model as an “artificial intelligence.” But it is misleading to the general public, who when they think of an artificial intelligence think of fictional characters like HAL 900 in 2001: A Space Odyssey, Skynet in The Terminator, Ava in Ex Machina, or Jarvis/Vision in the Marvel Cinematic Universe. LLMs are not people. They cannot think. They are not “intelligent” the way most people view it. They only “understand” the relationships between words as humans have used them.

Despite this, LLMs can do a surprisingly good job of mimicking human communication. To the non-specialist they certainly appear to be thinking. Indeed, some people have been tricked into believing an LLM is a person, and even have feelings for them. The same thing happened with ELIZA, the chatbot from the 1960s that merely picked out words from your prompts and repeated them back in a premade template. Imagine how people will be fooled by ChatGPT, which is incredibly more sophisticated?

You’ve probably heard hype about AI bringing about the apocalypse; or at least The Singularity, a moment in time when technology has advanced so drastically that the future becomes completely alien and unknowable. As for the apocalypse, some of these scenarios (the paperclip apocalypse, Roko’s Basilisk) are implausible to the point of being nonsensical. But there are imagined scenarios where AI becomes a danger to humanity, and we’ve seen the very beginning of this—automated policing and pre-trial risk assessment are obvious examples.

My wife and I are working on a science fiction novel where AI controls humanity, but it is based on a realistic scenario requiring two trends that are already happening in the real world: oligarchs employing AI to replace both physical and intellectual laborers and to control public opinion , and the public allowing this because AI provides ease and convenience. George Orwell thought the panopticon could only happen in a totalitarian state. He didn’t realize people would voluntarily sign on for it in exchange for two-day shipping.

Dammit, I Fumbled on My Chronoquill Roll: ChatGPT and TTRPG Game Design

But enough about AI. What about using it for your tabletop role-playing game?

I’ve been working on two RPG projects. The first is a Steampunk/Gaslamp Fantasy system-agnostic roleplaying world supplement based on a GURPS game I ran for a couple of years; I’m also writing short stories based in this world. The second is my ongoing Advanced Dungeons & Dragons 1st Edition campaign that takes place in a city based loosely on British Hong Kong. I decided I would try out ChatGPT and Midjourney as creative aids to working on these projects.

Here’s an example from the Steampunk game. I’ve been learning prompt engineering, the science and art of carefully wording AI prompts so that you get the kind of results you want. This prompt was very simple and brief, since I wanted to see what ChatGPT 4 would produce on its own.

I'm writing a Steampunk tabletop RPG. Define six main character stats and give them clever Victorian names.
Certainly! Here are six main character stats for your Steampunk tabletop RPG, along with clever Victorian names:
Cogitation (Intellect): Represents a character's analytical and problem-solving abilities, as well as their knowledge of scientific and mechanical principles.

Grit (Resilience): Reflects a character's physical and mental toughness, their ability to endure hardships, and their determination in the face of adversity.

Aetherflow (Energy): Measures a character's stamina, vitality, and ability to sustain prolonged physical or mental effort, resembling the flow of energy in aether-powered devices.

Invention (Creativity): Represents a character's imaginative thinking, inventiveness, and capacity to come up with innovative solutions to challenges.

Eloquence (Charisma): Reflects a character's persuasiveness, social charm, and ability to interact effectively with others in a world driven by social etiquette.

Chronoquill (Agility): Measures a character's speed, agility, reflexes, and dexterity, resembling the swift movements of a well-calibrated timepiece.
Feel free to adapt and expand upon these stats as needed for your tabletop RPG setting!

Now if you ask me, those aren’t too bad. Grit and Aetherflow overlap a bit too much, but that can be fixed. And the Victorianesque names are pretty good.

Now if you’re thinking, “ChatGPT certainly seems to understand what tabletop roleplaying character stats are like, and what ‘Steampunk’ means,” no, it doesn’t. This is based entirely on how words are used in relation to other words. But the illusion of intelligence and understanding is pretty striking.

I followed this prompt up with:

Now come up with ten character classes.
Of course! Here are ten character classes tailored to your Steampunk tabletop RPG:
Mechanist Engineer: Masters of mechanical devices and contraptions, Mechanist Engineers excel at crafting intricate gadgets and automatons to aid their allies and manipulate their foes.

Aethermancer: Harnessers of aetheric energies, Aethermancers can control and manipulate the mystical forces that power the Steampunk world, using them to cast powerful spells and disrupt opponents.

Steam Knight: Clad in steam-powered armor, Steam Knights are formidable melee combatants, wielding steam-driven weapons and using their heavy gear to withstand enemy attacks.

Clockwork Rogue: Masters of stealth and subterfuge, Clockwork Rogues use their mechanical expertise to create ingenious tools for lock-picking, sabotage, and espionage.

Skyship Captain: Skilled pilots of airships, Skyship Captains command the skies and lead their crew in epic aerial battles, using their knowledge of wind currents and steam propulsion to outmaneuver foes.

Alchemical Alchemist: Experts in the art of alchemy, Alchemical Alchemists concoct potent elixirs, volatile explosives, and mystical potions to provide both offensive and defensive support.

Tesla Artificer: Inspired by Nikola Tesla, these inventors specialize in harnessing electricity to create devastating ranged attacks and electrified gadgets.

Cogwheel Conjurer: Mages who channel their magic through ornate cogwheels and mechanisms, Cogwheel Conjurers can summon and control mechanical creatures and golems.

Victorian Explorer: Seasoned adventurers with a thirst for discovery, Victorian Explorers are skilled in survival, mapping, and unraveling the mysteries of uncharted territories.

Steambard: Charismatic performers who use their music and wit to inspire and motivate their allies, Steambards can also craft musical instruments and gadgets to enhance their abilities.
Feel free to modify, combine, or elaborate on these class concepts to fit the unique theme and mechanics of your Steampunk tabletop RPG!

Heh heh, “Alchemical Alchemist.” I’d use that just because it’s so dumb. Now some readers may see this list and think it’s pretty creative. Others will think it’s derivative, and contains no original ideas. My take is that there is some usable stuff here, as it adheres closely to Steampunk tropes.

But here’s the thing—if you’re trying to transcend established Steampunk tropes, you’re out of luck. ChatGPT can’t do that. You, the human, are still going to have to be actually creative.

In my experience working on my Steampunk game, ChatGPT couldn’t perform my creativity for me—the really cool stuff in my game world all came from my mind. But these responses gave me a jumping off point for my thinking, much as reading GURPS Steampunk and Castle Falkenstein gave me a jumping off point. I didn’t read these to steal ideas—they warned me what had already been done by others—but they inspired me, and so did the ChatGPT content.

But before we get to the pitfalls here, let’s talk about ChatGPT in my AD&D1e game.

AD&D1e Reloaded: Using ChatGPT to Spark Creativity

For my game, I’ve used ChatGPT for inspiration in a number of areas. I used it to create 1st Edition rules for Perception, Alchemists, and Tieflings; I used it to generate descriptions of locations; I even used it to design traps.

ChatGPT has read every main D&D resource, including each of the orange spine AD&D1e books. But this information has been compressed. So ChatGPT can explain how Armor Class works, but can’t tell you the specific AC of an elven fighter in leather armor with a buckler. It can describe how a spell works, but not list the specific component requirements or the range. It knows the general information, but not the numbers.

AD&D1e doesn’t have a Perception mechanic. I was trying to remember how Perception was handled back in the day, so I asked.

In 1e AD&D, how does Perception work?
In 1st Edition Advanced Dungeons & Dragons (AD&D), there isn't a unified "Perception" mechanic like in later editions. Instead, several different mechanics and systems were used to represent a character's ability to notice things, depending on the specific situation.
Listening at Doors: This is a basic check that many classes can make. Typically, a character has a 1 in 6 chance to hear something when they listen at a door, though this could vary based on the character's class.

Secret Doors: Elves, for example, had a passive ability to detect secret doors just by passing near them. If they actively searched, their chance to detect them increased.

Surprise: Some creatures or characters could surprise others. Typically, each side rolls a d6, and on a roll of 1 or 2, that side is surprised. Some classes or creatures might have modifiers to this roll.

Detecting Traps: Thieves (later called Rogues) had a percentage chance to detect traps. This mechanic used percentile dice, and as the thief leveled up, their chance to detect traps generally improved.

Other Checks: Some situations might call for specific checks or rolls to notice something. The details of these checks would depend on the Dungeon Master's discretion and the specific scenario.
It's important to note that much of 1e AD&D relied on the Dungeon Master's judgment and descriptions, so while there were mechanics in place for certain situations, many times it would come down to a conversation between the players and the DM.

See? General concepts, no numbers or details. In fact, if you ask for specific numbers, ChatGPT will just make them up. (When LLMs invent false facts, these are called “hallucinations.”) Look at this:

In AD&D1e, what is a third level thief's percentage chance to Detect Traps, based on information in the Player's Handbook?
In Advanced Dungeons & Dragons 1st Edition (AD&D 1e), a third level thief has a base percentage chance to Detect Traps of 25%. This information can be found in the Player's Handbook, specifically in the Thief section under the "Detect Traps" ability description.

Well, in the actual Player’s Handbook, the skill is called “Finding/Removing Traps” and a third level thief has a 30% chance, not 25%. So.

Devise 1e rules for Perception checks.
Certainly! Here's a simple way to incorporate Perception checks into 1e AD&D:
Ability Score Basis: Use the Wisdom (WIS) score as the basis for Perception, as it represents intuition, awareness, and common sense.

Basic Mechanic: When a character makes a Perception check, they roll a d20 and try to roll under their Wisdom score.

Modifiers:
Distance: For every 10 feet away an object or event is, impose a -1 penalty on the roll.
Low Light: If the area is dimly lit or obscured by fog, impose a -2 penalty.
Total Darkness: If the character is trying to perceive something in total darkness (and they don't have infravision or some similar ability), impose a -4 penalty.
Distraction: If the character is in a noisy or chaotic environment, impose a -3 penalty.
Racial Bonuses: Elves, with their keen senses, get a +2 bonus. Dwarves, with their underground instincts, get a +1 bonus when underground or dealing with stonework.

Class Bonuses:
Thieves: Due to their training, thieves get a +2 bonus when trying to perceive traps or hidden compartments.
Rangers: When in their favored terrain, rangers get a +3 bonus to Perception checks.

Using the Check:
Listening at Doors: Instead of the default 1 in 6 chance, a character can make a Perception check. Use the modifiers above to adjust for the thickness of the door or any noise on the other side.
Secret Doors: An active search allows a Perception check with racial bonuses applied.
Detecting Traps: Thieves can opt to use their specific trap-finding skill or make a general Perception check if they're looking for something less specific.
Natural 1 and Natural 20: If you like the idea of critical successes and failures, a roll of 1 on the d20 could be an automatic success, and a roll of 20 could be an automatic failure, regardless of modifiers.
Remember, the beauty of AD&D is its flexibility. These rules are just a guideline, and the Dungeon Master should feel free to adjust or expand upon them to better fit their campaign or player group.

Look, it even made up specific numbers! Which is fine, in this instance where the numbers don’t already exist.

So is this usable? It’s not bad! Obviously, ChatGPT is mining future editions for ideas, which it did when I asked it for rules about Alchemists and Tieflings. Some of the overlap with existing Thief skills is unnecessary (and has the side effect of making a high Wisdom invaluable to Thieves). The truth is, I could write a whole blog post just about this topic, Perception in 1e. In fact, I plan to.

Anyway, a human DM has to look at this critically. You can’t take rules that ChatGPT generates and just throw them into your game. You have to be thoughtful. But it’s a good place to start!

Asking ChatGPT to generate the crunchy stuff worked fairly well. Asking it to be creative, less so. For instance, I would describe a location to ChatGPT, and it would provide a description.

The shrine stands as a testament to the enigmatic nature of Chokyeong the Reputable. Made of dark, polished wood, the structure is modest in size but intricate in design. The entrance is adorned with crescent moon motifs and intricate patterns that seem to shift and change when viewed from different angles, a nod to the trickster god's unpredictable nature.
Inside, the air is thick with the scent of sandalwood incense. The walls are lined with small offerings from devotees: tiny silver coins, feathers, and mysterious trinkets. At the center stands a large icon of Chokyeong in his wolf form. The wolf, made of moonstone and obsidian, has eyes that seem to gleam with an inner light, giving it an almost lifelike quality. Its posture is both regal and playful, capturing the essence of the trickster god.
Behind the icon, a tapestry depicts the various phases of the moon, with Chokyeong's wolf form playfully chasing and being chased by the celestial body, symbolizing the eternal dance between mischief and time.
The shrine, though small, exudes an aura of reverence and mystery. It's a place where the boundaries between the mortal realm and the divine seem blurred, and where the laughter of the trickster god might just be heard on the wind.

Not bad, right? My players enjoyed these descriptions, for a while. But over time, the text provided by the LLM ended up being too long and too florid for the players. I can, of course, edit these descriptions down, and I do. But the truth is, I’m still inventing all the general concepts for these locations. If I’m lazy or in a hurry, I can go with the generated descriptions. But since I’m doing half the work, I might as well write the rest myself.

One last thing before I get to my conclusion. ChatGPT did one thing amazingly—well, badly, but in an amazing way. I create reading for my players, usually in-character newspapers with worldbuilding details and adventure prompts. I decided for this game, I wanted my newspaper copy to sound “medieval.”

I wrote this in everyday English, and then asked ChatGPT to translate it into Early Modern English (Shakespeare’s English):

NO NEED FOR WORRY
Princess Lagertha Assureth the City’s Safety from Chosun Invaders

The Imperial Crown Princess Lagertha Hammerhand, speaking extemporaneously at a lordly feast, did assure those present that Lord Gath, the Colonial Governor, purposeth to defend the City ‘gainst any threat of invasion by the forces of the Sevennames Empire.

Lady Lagertha, the wondrous ten-year-old daughter and heir to Ingvones Emperor Ledo XIV Hammerhand, proclaimeth her father’s full support to Lord Gath and the Imperial Navy ‘gainst the Sevennames Emperor’s feigned right to the Crown Colony.

‘He is not the rightful Emperor,’ quoth the Princess, speaking of the Sevennames leader Kim Sung-Heon. ‘He’s but an usurper. Our treaty was with the Namgung Emperor. The City of the Transcendent Moon shall stand as a Crown Colony whilst my father yet breathes.’

Now that’s not academically-correct Early Modern English. But it’s fucking hilarious. I also got ChatGPT to translate text into a florid “East Asian” style.

Ah, let me regale you with the tale of Nikona, a realm of enchantment nestled in the sea east of Zhongguo, where ancient myths dance amidst the echoes of history. The main island, known as Yomoru, boasts the grandeur of the Serpent's Spine Range, a mountainous expanse that cleaves the island from north to south, a testament to the might of the natural world. Within this cleft, the lifeblood of Nikona, the Ginryūgawa River, courses with vitality, nurturing both Nikonan people and spirits alike, who dwell within the vertical cities and towns etched into the cleft's resplendent walls.

Remember none of these ideas came from ChatGPT. I wrote detailed copy, then it was translated into a writing style. The new style isn’t “better,” but it is more fun.

Wait, Is That All You’ve Got?: The Verdict on ChatGPT

So I’ve been using ChatGPT-enhanced ideas in my AD&D1e game. How has it gone? Some of the “enhancements” have gone over well, some haven’t. But really, it comes down to one problem.

When I let ChatGPT come up with ideas, the players see them coming a mile away.

For one example, I described a vault puzzle to ChatGPT and asked it to come up with a description and some possible solutions. I picked the solution that seemed most creative. Well, my thief player guessed it immediately.

In another instance, I had ChatGPT design a novel monster, with stats, specifically designed for this party. The fighter one-shotted it on the first round.

So, that’s the end of my career as a lazy DM. I’ll still use ChatGPT to create entertaining writing styles or to inspire ideas, but the fact is the hard creative work is up to me, just as it always was. And honestly, that’s better.

You’re Supposed to Play a Thief, Not Actually Be One: ChatGPT and Copyright

Before I conclude, I want to say a word about ChatGPT and the theft of ideas. As I said, ChatGPT mined the Internet and digital documents for everything it knows. That means its output is based on the work of countless writers whose permission to scan their work was never even considered, much less requested.

So does ChatGPT respond with stolen ideas? As far as I can tell, not directly. For a number of “original” (that is, unfamiliar to me) ideas ChatGPT gave me, I did some internet searching to see if I could locate a source. I didn’t find any. Of course, very many digital documents aren’t easily searchable, and aren’t scanned by search engines.

What ChatGPT seems to do is examine a topic and then spit out what I’d call “clever pablum.” Clever because it might seem creative; pablum because it’s really just warmed-over tropes. That’s why my players saw it coming. They’ve seen it all before, just packaged differently.

Just as Wikipedia has an admitted bias of “neutrality,” which in practice means adhering to conventional wisdom and editing out new, original, or fringe ideas, ChatGPT seems to have an unspoken bias of what I can only call mediocrity. Americans think of mediocrity as being “bad,” but it just means average or ordinary. Ask for originality, and you’re going to get established tropes, because that’s the “average” information in its dataset.

Anyway, that’s my experience with ChatGPT and TTRPGs. I was going to get into creating images with Midjourney, but I think that’s another blog post.

ChatGPT, give me a one paragraph synopsis of this blog post.
In this blog post, the author reflects on their experience using ChatGPT, a large language model (LLM), as a tool for tabletop role-playing game (TTRPG) design. Initially motivated by concerns about AI potentially replacing their role as a business writer, the author explores ChatGPT's capabilities and limitations. The post covers various aspects, such as the nature of LLMs, their ability to mimic human communication, and the author's experiments in using ChatGPT for creating content in a steampunk TTRPG setting and an Advanced Dungeons & Dragons (AD&D) 1st Edition campaign. The author highlights both the potential benefits and pitfalls of incorporating AI-generated content in TTRPG design, emphasizing the need for human creativity and critical evaluation. The post concludes with the author's verdict on ChatGPT and a brief mention of ethical considerations related to the potential use of AI-generated ideas without proper attribution.

Leave a Reply

Your email address will not be published. Required fields are marked *