Can ChatGPT generate D&D rules? As D&D freelancer, that question is top of mind these days.
Not “can ChatGPT generate better rules than a good game designer?” I think the answer to that is quite obviously no. But “can ChatGPT generate content good enough that a publisher might think, hmm, paying 0/word is better than paying 5, 10, 15 cents per word?”
Let’s see where we stand. I’ve just made a ChatGPT account and I’m going to give it a few prompts to see whether it’s good enough to take my job – or work for me, if I’m inclined to hire it! Fair warning, I’m rooting against ChatGPT, but I will try to be fair here.
Task 1: Translation. I’ll start off with an easy one just to calibrate my expectations. I’ll assign a simple, right-or-wrong task that anyone could do with a little research.
In the 2014 Players Handbook, one of the images has some runic letters on it, and the the letters spell out a supremely disappointing message: in something akin to Elder Futhark the runes spell something like “This is the Photoshop Lorem Ipsum Text.” A huge wasted opportunity for a hidden Easter egg. This could have been a hint about an upcoming product or a little “no x were harmed” joke! Could we use ChatGPT to come up with some more plausible text to jam into this image?
To see if ChatGPT can do this, let’s give it a nice easy question. My prompt:
Transliterate the first sentence of pride and prejudice into elder futhark
Elder Futhark is a runic alphabet with a wikipedia page and lots of information about it on the web, and Pride and Prejudice is a public domain book with a supremely famous first sentence: “It is a truth universally acknowledged, that a single man in possession of a good fortune must be in want of a wife.” This should be an easy one.
ChatGPT’s response:
ᛁᛏ ᛁᛋ ᚫ ᛏᚱᚢᚠᛋᚫᚫ ᚢᚾᛁᚱᛡᛋᛚ᛫ ᚫᚦ ᚫ ᛋᛁᚾᚷᛚᛖ ᛗᚢᛋᛏ ᛒᛖ ᛁᚾ ᚹᚫᚾᛏ ᚩᚠ ᚫ ᚷᚢᛏ ᚠᛟᚱᚢᚾᛏ, ᛗᚢᛋᛏ ᛒᛖ ᛁᚾ ᚹᚫᚾᛏ ᛟᚠ ᚫ ᚷᚮᛞ ᚠᚮᚱᚢᚾᛏᛖ, ᛗᚢᛋᛏ ᛒᛖ ᛁᚾ ᚹᚫᚾᛏ ᚢᚾᛁᚱᛡᛋᛚᛖ.
This translates to
it is a trufsaa unirhsl ath a single must be in want of a gut forunt, must be in want of a g?d f?runte, must be in want unirhsle.
Arrant nonsense! The response is not all Elder Futhark runes, and it’s certainly not all Jane Austen. While it has stretches of being plausible, it gets into a sad loop at the end. Simple transliteration should be an easy task for a computer, so I’m astonished that ChatGPT failed so hard right out of the gate.
Task 2: Monster Creation. I write a lot of D&D monsters, so let’s see how ChatGPT copes with that task.
Prompt: write a D&D stat block for a CR 20 monster suitable for a campaign villain.
Here’s what ChatGPT came up with:
Eldrith the Malevolent
Large Fiend (Demon), Chaotic EvilArmor Class: 19 (Natural Armor)
Hit Points: 270 (20d12 + 140)
Speed: 40 ft., Fly 60 ft.STR 24 (+7)
DEX 18 (+4)
CON 24 (+7)
INT 20 (+5)
WIS 20 (+5)
CHA 22 (+6)Saving Throws: Str +14, Dex +11, Con +14, Int +12, Wis +12, Cha +13
Skills: Arcana +12, Deception +13, Insight +12, Perception +12
Damage Resistances: Cold, Fire, Lightning
Damage Immunities: Poison; Bludgeoning, Piercing, and Slashing from Nonmagical Weapons
Condition Immunities: Charmed, Frightened, Poisoned
Senses: Truesight 120 ft., Darkvision 120 ft., Passive Perception 22
Languages: Abyssal, Common, Infernal, telepathy 120 ft.
Challenge: 20 (25,000 XP)Legendary Resistance (3/Day). If Eldrith fails a saving throw, she can choose to succeed instead.
Magic Resistance. Eldrith has advantage on saving throws against spells and other magical effects.
Innate Spellcasting. Eldrith’s spellcasting ability is Charisma (spell save DC 21). She can innately cast the following spells, requiring no material components:
At will: detect magic, dispel magic, telekinesis, plane shift
3/day each: feeblemind, imprisonment, teleportSpellcasting. Eldrith is a 20th-level spellcaster. Her spellcasting ability is Intelligence (spell save DC 20, +12 to hit with spell attacks). She can cast spells from the sorcerer and warlock spell lists. Eldrith has the following sorcerer and warlock spells prepared:
Cantrips (at will): eldritch blast, mage hand, prestidigitation, minor illusion
1st level (4 slots): eldritch blast, mage armor
2nd level (3 slots): misty step, suggestion
3rd level (3 slots): counterspell, dispel magic, fireball
4th level (3 slots): dimension door, greater invisibility
5th level (3 slots): teleport, wall of force
6th level (2 slots): chain lightning, globe of invulnerability
7th level (2 slots): finger of death, plane shift
8th level (1 slot): dominate monster, feeblemind
9th level (1 slot): time stopActions
Multiattack. Eldrith can use her Frightful Presence. She then makes three attacks: one with her bite and two with her claws.
Bite. Melee Weapon Attack: +14 to hit, reach 10 ft., one target. Hit: 19 (3d8 + 7) piercing damage plus 14 (4d6) poison damage.
Claws. Melee Weapon Attack: +14 to hit, reach 5 ft., one target. Hit: 15 (2d6 + 7) slashing damage.
Frightful Presence. Each creature of Eldrith’s choice within 120 feet of her and aware of her must succeed on a DC 21 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 Eldrith’s Frightful Presence for the next 24 hours.
Legendary Actions
Eldrith 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. Eldrith regains spent legendary actions at the start of her turn.
Attack. Eldrith makes one claw attack.
Spellcasting (Costs 2 Actions). Eldrith casts a spell from her list of prepared spells, using a spell slot as normal.
Teleport. Eldrith magically teleports, along with any equipment she is wearing or carrying, up to 120 feet to an unoccupied space she can see.
Lair Actions
On initiative count 20 (losing initiative ties), Eldrith takes a lair action to cause one of the following effects; Eldrith can’t use the same effect two rounds in a row:
Eldrith casts teleport to switch places with another creature within her lair.
Eldrith casts wall of force to create a barrier, dividing the battlefield.Regional Effects
The region containing Eldrith’s lair is warped by her magic, creating one or more of the following effects:
Eldrith’s presence causes plants to wither and die within 1 mile of her lair.
Creatures within 1 mile of the lair are subject to unsettling dreams and fitful sleep, causing them to suffer from exhaustion.
Non-magical fires within 1 mile of the lair burn with an eerie, green flame.These effects fade if Eldrith is slain.
Description: Eldrith the Malevolent is a fiendish being of immense power, with command over both sorcery and warlock magic. Her lair is a dark and twisted realm, corrupted by her malevolent presence. Eldrith is a cunning and manipulative adversary, using her vast array of spells and abilities to outsmart and overpower her foes. She is known for her ability to strike fear into the hearts of even the bravest heroes, and her cruel and sadistic nature makes her a formidable foe in any battle.
Roleplaying Notes: Eldrith is a master of deception and manipulation, and she delights in toying with the emotions and minds of those who oppose her. She is relentless in her pursuit of power and is willing to do whatever it takes to achieve her goals. Eldrith is fiercely intelligent and strategic, always thinking several steps ahead of her enemies. She revels in chaos and revels in the suffering of others. Eldrith is a truly malevolent force to be reckoned with, and facing her in battle will be a challenge that only the most skilled and determined adventurers can hope to overcome.
OK, there’s a lot here, both good and bad. On the good side, ChatGPT got the format approximately correct. It realized that a campaign villain should be a legendary monster (on the other hand, since nearly all CR 20+ monsters are legendary, maybe that’s a coincidence based on the way the dataset is weighted?) Hit points and other stats are plausible for CR 20. It seems to have correctly calculated skills and saving throws, which is amazing for a bot that mangled a single line of Pride and Prejudice beyond all recognition. It made a few errors, of course. D12 is the wrong die type for a Large monster, though it got the hit point caculation right otherwise. It has eldritch blast listed as both a cantrip and 1st-level spell; I guess someone named Eldrith really has to lean into eldritch blast. But this is just nitpicking. The real problem with Eldrith is that she’s boring and her mechanics don’t support her story.
Mechanically, Eldrith is a reskin of every 5e dragon ever, which is unfortunate since we’re all sick of stock 5e dragons at this point. Frightful Presence followed by a claw/claw/bite is marginally plausible for a fiend, I guess, but we all know what we’re looking at here, and that’s a dragon without a breath weapon. The spellcasting tucked in the legendary actions is interesting: I don’t mind it, to tell the truth. But I’m not sure it will salvage the boring gameplay of this dragon sans breath weapon. Let’s take a closer look.
What spells will Eldrith cast? Time stop is good but she doesn’t have a lot of buff spells to back it up (mage armor is useless given her natural armor; greater invisibility, wall of force, and globe of invulnerability all use concentration). Dominate monster and suggestion are not great in combat. Plane shift, misty step, dimension door, teleport, her bizarre extra innate teleport, and the teleport legendary action are amazingly redundant (maybe not so amazing considering the results of the Jane Austen task). So what spells will this campaign boss cast? After using feeblemind, chain lightning, and finger of death, Eldreth is maybe best off using her high-level slots to upcast fireball in order to make up for the breath weapon she’s missing. Overall, she’s got a somewhat random assortment of spells, best suited to running away (not a bad skillset for a campaign villain tbh).
How well does Eldreth fare lorewise? Her description is cliche-ridden (cunning and manipulative adversary, cruel and sadistic, master of deception and manipulation, strikes fear, delights in toying with the emotions and minds of those who oppose her, relentless in her pursuit of power, got it, yup, yup.) And the description doesn’t particularly match her mechanics. She has no unique stat block mechanics that let her be cunning and manipulative — or unique stat block mechanics of any other kind. Spell-wise, suggestion and dominate monster are the best matches for her flavor. I’d really like to see more charm, fear, phantasmal force, and modify memory on her spell list. And she desperately needs a signature ability — something, anything, unique, preferably something that matches her manipulative-mastermind lore.
What strikes me about Eldreth is that she is unfinished. She’s a half-baked cake with no icing. ChatGPT did the easy part for us, but left the hardest part – “make it interesting” – undone.
Task 3: Magic Item Enchantment. OK, maybe a 20th level monster was a bit of a stiff challenge. Let’s try something significantly easier: a magic item.
My prompt:
Write a magic item description in the style of the Fifth Edition D&D Dungeon Master’s Guide. The magic item should be named “The Rose of Estes” and should provide specific benefits that interact with D&D game rules. Word count: 100-150 words
ChatGPT’s response:
The Rose of Estes
Wondrous Item (Amulet), Legendary (Requires Attunement)
This exquisite amulet holds a crimson rose within a crystalline pendant. Crafted by the enchanter Estes, it embodies the eternal cycle of nature. While attuned, you gain the following:
Blossom’s Vigor: As a reaction when hit, the rose blooms, granting you temporary hit points equal to your proficiency bonus.
Photosynthesis Affinity: When exposed to sunlight, you regain hit points equal to half your level at the start of your turn.
Verdant Resilience: You have advantage on saving throws against poisons, diseases, and effects that cause exhaustion.
Curse: You are compelled to protect plant life. If a plant is harmed within 30 feet, you must make a Wisdom saving throw (DC 15) or take 1d6 psychic damage.
Roleplaying Note: The Rose pulses with the rhythm of nature. Its petals shimmer when danger nears, a testament to Estes’ love for the wild. When worn, it’s as if the amulet breathes in time with the natural world.
Shoot, that’s … not bad. The lore is maybe a bit bland, but I like the curse. It is, however, somewhat OP. Too OP? Temporary hit points AND regeneration every turn is pretty strong. Let’s compare other items. Ring of Regeneration (very rare) grants regeneration of 3 hp per 10 minutes, and the Ioun stone of regeneration (legendary) grants 15 hp per hour. Even the Eye and Hand of Vecna combined — two artifacts — only grant a regeneration factor of 1d10 hp per turn. Meanwhile, the Rose of Estes grants your level HP per turn (probably 15 or more hp considering the item’s rarity), though it’s limited to sunlight situations, which is a significant limitation. I wonder, though, is this a case of the Rose of Estes being OP or of the DMG items being too conservative? When characters are high-level enough to use legendary items, between-battle healing is not exactly hard to come by.
Balance aside, the Rose of Estes description misses on the most important qualification for freelance writing. It’s 160 words — higher than my very clearly-spelled-out word count. Making it… just like every freelance writer turnover ever.
My conclusion: ChatGPT is sometimes astonishing in what it can do. It’s amazing that a bot can intuit and regurgitate game rules! But despite that, it’s just not there yet as a RPG-writing tool. Half the time, it gives you solid but uninspired work. That might be useful in a certain niche, but you can’t trust it because the other half of the time, it spouts nonsense with the appearance of utter confidence. Just as AI art prompts so often produce conventionally attractive folks festooned with extra fingers and other tendrils, ChatGPT’s D&D rules prompts generate a mixture of the banal and the broken.
Right now, ChatGPT is not a finished product but a promise — a promise to democratize art by separating creative workers from their means of production. Inasmuch as I just want to watch the world burn, I’m curious to see where it will go next. As AI improves, we’ll have to figure out a way to live alongside it. After all, you can’t get the genie back in the bottle. All you can do is find a way to survive the genie’s Frightful Presence and claw/claw/bite.
I’ve had better success when I use ChatGPT in a conversational mode. I present a prompt, it gives a response, which clarifies my thinking or occasionally sparks an insight, I prompt with a refinement, and the cycle continues.
It takes a bit of practice to settle into that mode, to avoid hoping I’ll get something wonderful and complete in response to a single perfectly crafted prompt.
Here’s me developing a haunted mansion adventure. Skip ahead to Blackthorn Manor.
https://chat.openai.com/share/2c8c77c8-b1bf-4e22-9cb2-07556461445d
I think it’s worth noting that you were using the free GPT-3.5 model, which is several years old. This makes sense, given the proposed “budget” of $0/word, but the more recent GPT-4 model you can use with a paid account is noticeably “smarter” and more powerful, as well as having access to plugins (e.g. you can give it web search, a calculator etc.)
While GPT-4 isn’t free (and I don’t have a subscription), it *is* cheap enough that Microsoft are willing to pay for a version of it to be included for free in Bing. Pasting your first question into Bing, it searched for a table of Futhark letter equivalents and then gave me this:
ᛁᛏ ᛁᛋ ᚨ ᛏᚱᚢᚦ ᚢᚾᛁᚹᛖᚱᛋᚨᛚᛇ ᚨᚲᚲᚾᛟᚹᛚᛖᛞᛄᛖᛞ, ᚦᚨᛏ ᚨ ᛋᛁᚾᚷᛖᛚ ᛗᚨᚾ
ᛁᚾ ᛈᛟᛋᛋᛖᛋᛋᛁᛟᚾ ᛟᚠ ᚨ ᚷ
Of course, this is a toy example more easily accomplished by any number of free websites, but using one of those free web apps to check, it appears to be correct (although it cuts off part-way through a word at the end, and typos “single” as “singel”.) I’m not sure if the ability to search the web or the more powerful model was decisive here. Still, a good example of how fast the state of the art is moving – just because GPT-3.5 can’t do something doesn’t mean GPT-4 necessarily can’t.
GPT-4 is *not*, however, obviously a much better writer than GPT-3.5 just off the cuff (IMO). I do think it probably has more *potential* for good writing, but both models generally need some pushing to get into the right “headspace” for even moderately well-written material. I think you (accidentally?) did that with your Rose example; it’s an evocative prompt that makes the AI more likely to predict something cool will come after it than average.
A final point worth noting: it can take feedback and criticism. If you ask for a more evocative version of the boss or a more balanced version of the amulet, you’ll probably get one. Enough repeated rounds of feedback to get something actually *good* can often take more time and energy than just writing it yourself, however!
Hey Paul,
I think you are write that the genie is not going back in the bottle. You will find you get rather better results with GPT 4, and it might be worth $20/month to experiment with.
I also find you have to give chat gpt a bit more to work with in the prompt to get good results. One trick is to ask it to give you several options, and then you pick out what sounds good to explore further.
For example, instead of your boss monster query, I wrote this in GPT 4:
please suggest 10 amazing and original D&D boss monsters suitable for a campaign climax. give me one line of description each
And here are the results
1. Abyssal Siren: A hauntingly beautiful mermaid with a choir of cursed souls, whose song can bend reality and manipulate emotions.
2. Celestial Chimera: An amalgamation of various celestial creatures, it guards a portal to the heavens, and its different parts have unique powers of the gods.
3. Ebonshade Lich: A master of shadows and time, capable of casting a region into eternal night and summoning twisted versions of heroes from alternate timelines.
4. Elemental Monarch: Born from the fusion of all elemental planes, it shifts form between earth, fire, wind, and water, each phase with distinct powers.
5. Gloomweaver Spider Queen: A massive arachnid queen from the Shadowfell, she weaves reality-distorting webs and commands legions of shadow spiders.
6. Nexus Golem: A massive construct made from fragments of destroyed planes, drawing abilities from each plane and constantly shifting its form.
7. Pandemonium Pharaoh: An undead king from an ancient civilization, wielding chaotic magic that reshapes the battlefield and controls enemies’ senses.
8. Temporal Serpent: An ethereal snake encircling a massive hourglass, with the power to age, rewind, or freeze foes and parts of the battlefield in time.
9. Verdant Behemoth: A gargantuan living forest creature, with the ability to manipulate nature, regrow, and summon woodland allies instantly.
10. Voidmind Kraken: A cosmic cephalopod existing partially in another dimension, its tentacles pull foes into void pockets, and its gaze drives men mad.
I then asked it to stat out the Nexus Golem:
Nexus Golem
Gargantuan Construct, Neutral
Armor Class 20 (natural armor)
Hit Points 375 (20d20 + 140)
Speed 30 ft.
STR 30 (+10)
DEX 10 (+0)
CON 24 (+7)
INT 3 (-4)
WIS 11 (+0)
CHA 1 (-5)
Saving Throws STR +16, CON +13
Damage Resistances bludgeoning, piercing, and slashing from nonmagical attacks
Damage Immunities poison; psychic
Condition Immunities charmed, exhaustion, frightened, paralyzed, petrified, poisoned
Senses darkvision 120 ft., passive Perception 10
Languages understands the languages of its creator but can’t speak
Challenge 20 (25,000 XP)
Special Traits:
Multiplanar Being. The Nexus Golem exists across several planes simultaneously. It has advantage on saving throws against spells and other magical effects. Additionally, when the Nexus Golem is targeted by a spell or ability that would banish or transport it to another plane, it can choose to negate that effect.
Planar Absorption. When a spell of 1st level or higher targets the Nexus Golem, it gains a number of temporary hit points equal to twice the spell’s level and can use a reaction to make a slam attack against a creature within its reach.
Actions:
Multiattack. The Nexus Golem makes two slam attacks.
Slam.
Melee Weapon Attack: +16 to hit, reach 10 ft., one target.
Hit: 32 (4d10 + 10) bludgeoning damage.
Planar Surge (Recharge 5-6).
The Nexus Golem releases stored planar energy in a 30-foot radius. Each creature in that area must make a DC 18 Dexterity saving throw, taking 45 (10d8) force damage on a failed save, or half as much damage on a successful one. Additionally, the Nexus Golem can choose one of the following effects to impose on those who failed their saves:
Elemental Burst: Targets are scorched with elemental energy, taking an additional 18 (4d8) fire damage.
Temporal Distortion: Targets are aged or reverted in time, becoming either one age category older or younger (player’s choice) until the end of their next turn.
Abyssal Grasp: Shadowy tendrils from the Shadowfell grasp onto targets, restraining them until the end of their next turn.
Reactions:
Shift Plane. When the Nexus Golem takes damage, it can use its reaction to shift slightly between planes, granting it resistance to that damage type (if it doesn’t already have it) until the end of its next turn.
I think this a reasonable result that could be “polished up” into something useful. I feel like the future of these tools is as enhancements to human effort rather than simple creative replacements, but perhaps I’m naive. As you say, they are not going away.
I see that others have already written comments similar to what I was going to suggest– to take advantage of the “chat” aspect of ChatGPT. I find that I get pretty good results when I treat ChatGPT as a well meaning but junior assistant.
For example, with the monster created in Task 2, I suggest responding with something like, “Thanks. That’s a good start but I feel like she desperately needs a signature ability — something, anything, unique, preferably something that matches her manipulative-mastermind lore. Please suggest three options for a signature ability.”
With the magic item description you can tell ChatGPT something like, “I think the word count is incorrect. Please review and edit your description such that it uses between 100 and 150 words.”
I certainly agree that ChatGPT isn’t an “automatic ‘I win’ button” but it is certainly a good tool for a DM. Like having a friend or assistant that you can bounce ideas around with when you are creating things.
Yuck. A publisher pulling this crap would inspire me to leave their product on the shelf.
It is incredible how fast this technology is developing but as a DM I dont need or want it in my DnD.
You want to use AI, cool. Go cure cancer. Engineer a better everything, have at it, but leave art alone. That’s my opinion.
lmao, any publisher that used any AI-generated trash, whether writing or art, would get an immediate and permanent boycott from me. That’s how we put the genie “back in the bottle,” by utterly sanctioning it to hell and making it the touch of death for any business that uses it to put people out of a job in the name of greed.
And don’t be deluded, the only reason to involve “generative AI” in a published creative product, in any form of media, is greed. There is no excuse.
Your blog is appreciated for its usefulness and informative content.
IFDA is the best institute for Computer Courses in Delhi and after that any courses completed IFDA is providing paid internship & 100% job placement.
Best Computer Course in Delhi
Having worked with many freelance game writers during my career, I’m sad to say that those examples stand up surprisingly well compared to what humans often produce under stress from deadline pressure.