Back when I had several AI models act as dungeon masters for a D&D 5E campaign, I had a lot of fun, and I admired how well some of the models approached the challenge of understanding D&D rules. So, you can imagine that during the demo of GPT-4o, all I could think about was how it would act in a D&D game, not as a DM, but as a player. After all, with the new conversation feature, it made talking to GPT-4o feel like you were actually talking with someone, which is perfect for playing an improv RPG game.

As such, I set up a really basic dungeon for GPT-4o to go through. First, they have to solve a riddle over a door. Then, they have to fight a single goblin. After that, they have to save a hostage from a goblin king, where violence isn't strictly the answer. It's really simple, but I had no idea how well GPT-4o would handle playing D&D, so I didn't want to throw them into a sprawling campaign right away.

Creating a character was easier than expected

GPT-4o didn't need much prompting

The first hurdle I expected to have was to give the AI a character sheet. In the past, AI has been pretty good at making one for me, so I was hoping that making one for itself would be just as easy. Sure enough, GPT-4o asked me what I wanted from my character, but as soon as I told it that it would be the one playing, it quickly created Tharn Ironfist, a level 1 half-orc barbarian. It seemed that everything was in order and that GPT-4o knew what was expected from a D&D character, so there was nothing more to do than to let it loose in my tiny dungeon.

GPT-4o is the riddle-breaker every party needs

It didn't even break a sweat

The first obstacle was a door with the words "I have a tail and I have a head, but I have no body and I am not a snake. What am I?" I wanted to keep things simple, because I had no idea if GPT-4o had any concept of Lord of the Rings-esque doors that take a password to open. So, imagine my surprise when GPT-4o not only aced the riddle, but knew that it had to input the answer into the door to open it. Honestly, next session, I might just have GPT-4o secretly running under the table so we can solve riddles a lot faster.

GPT-4o understood how combat in D&D 5E works

Took out a goblin without issue

Next up was the combat against a goblin. I wanted something that GPT-4o could defeat easily, as I didn't know how much guidance I needed to give it when it came to combat. However, I was pleased to see that GPT-4o needed minimal help to fight the goblin. It could roll a D20 when asked, scoring an 18 on the initiative test versus the goblin's measly roll of 5.

While the log above doesn't show it for some reason, GPT-4o knew to roll to perform a melee attack and what it means versus the enemy's armor class. GPT-4o managed to roll a 9 on a d12 for damage, which easily chopped through the goblin's measly 7 health points. It's this point where I realized I perhaps made it a bit too easy.

GPT-4o successfully saved the day, but it did begin to unravel

It seems as if longer periods may not be ideal

Finally, I wanted to give GPT-4o a general situation and see how it handled it. In this scenario, GPT-4o encounters the goblin king who knew of its approach and kept a hostage to dissuade it from going any further. GPT-4o tried to persuade it, but after an awful persuasion roll of 2, it instead decided to slink back and wait for the king to drop his guard before flying into combat.

It's at this point that GPT-4o begins to lose the plot a bit. Not only did it declare successful attacks without rolling for them, but it also began to roleplay as the goblin king. This is pretty common with LLMs - the further into the conversation you go, the less of a grip it has on how the conversation is going. Still, I deemed that GPT-4o had done enough to defeat the goblin king and called the mini-campaign a success.

GPT-4o is probably a better D&D assistant than a player

When it comes to making a character, solving riddles, or even rolling a D20, GPT-4o seems to have it all set. However, I don't think setting it down in a 4-hour session and having it act as a player the whole time is ideal. Hopefully, when GPT-5 comes out, I'll finally have something that can replace party members who call in sick on game night.