• The VOIDRUNNER'S CODEX is coming! Explore new worlds, fight oppressive empires, fend off fearsome aliens, and wield deadly psionics with this comprehensive boxed set expansion for 5E and A5E!

RPG Evolution: That Time an AI Played RPGs

Here's how AI works as a player and DM.

Artificial Intelligence is being applied to everything, including role-playing games. Here's how it works as a player and DM.


Picture courtesy of Pixabay.

Your Move AI​

It's perhaps not surprising that one of the ways an AI can be tested is how it performs within the artificial environment of a role-playing game. In theory, "anything can be attempted," which is a good test of an AI's capabilities. Without the usual guardrails, can an AI navigate a RPG world? And if it can do so convincingly, can it create a world in reaction to players? We're about to find out.

Seeing the LIGHT​

Researchers from the Georgia Institute of Technology in Atlanta and Facebook AI Research combined techniques to test their AI's approach. They used a text-based multiplayer game called LIGHT, a text-based fantasy RPG developed by Facebook to study communication between human and AI players. The game is populated by crowdsourced objects, characters, and locations, all described via text.

With the framework in place, the team gave the AI 7,500 crowdsourced quests and armed it with a knowledge graph (a database of subject-verb/object agreements) so it could navigate the world. With a plan of action and tools to achieve its goal, the AI was set loose in LIGHT.

At times the AI, who in one simulation (role-playing as a witch) had a cat familiar named Helix, threatened a fairy by demanding a bucket or "I'll feed you to my cat!" When the AI role-played a dragon, it was tasked with specific missions like hoarding gold. And it did eventually get there, even if there were some hiccups along the way.

If an AI could navigate a RPG successfully, the next question is if it could run one. Enter AI Dungeon.

Roll for Initiative​

AI Dungeon was first conceived of and designed by programmer Nick Walton during a hackathon event in 2019. It has since been developed by his company, Latitude. AI Dungeon generates stories for single- and multi-player groups in a variety of genres via text descriptions. It spits out narrative text in the second-person, similar to Multi-User Dungeons and Interactive Fiction.

There are flaws however, and these flaws are a good overall representation of the risks of AI in general. It essentially uses a library of phrases and parrots them back, which sometimes doesn't always make sense.

In my playtest, my ranger Talien killed a mythical beast that "looked like a giant wolf but about the size of a bear." After killing it and inspecting the corpse, the AI Dungeon declared I had "killed a wyvern" and that the corpse might be worth something. I tried to extract poison from the wyvern (assuming it was a D&D-style wyvern), and the AI Dungeon responded that I injected poison into the beast instead. Randomly, my employer shouted (from where?), "Are you ready? Go!"

Unlike text-based games that look for specific prompts, AI Dungeon allows anything to be attempted in true RPG fashion. You can type anything and the AI types back in an attempt to flex to what you're doing. It wasn't long before my employer ("a noble man"), asked me to track down Baron Joridal, a tall man with a large mustache and a small scar on his right cheek.

The downside of "anything can be attempted" means that the game can stray into darker territory. Like most of AI, the AI reflects back at players what it learns, and often that content includes material inappropriate for children. AI Dungeon has some safeguards in place to address those concerns, but it's not perfect.

The Future​

For the moment, AI isn't going to replace players or DMs any time soon. But as the datasets get better, the AI gets faster, and humans help it refine responses, we may get AI that's indistinguishable from the rest of us soon.

log in or register to remove this ad

Michael Tresca

Michael Tresca

Actually, I have an interesting tidbit or you:
I took the AI at one point, in order to brainstorm a new setting from the ground up. I advised it to create three personas and name them. We would then have a brainstorming session to develop said setting.

It ran without me. Those three simulated people discussing and adding to a setting they devised. In the end, it pretty much looked like most of the brainstorming sessions I've been in, and I've been in a bunch (games industry for two decades).

So what happens when you try to replace the players with an AI? <looks at players everywhere> Shape up. You can be replaced too.

DnD is a complex game. Having a AI play or DM DnD would be an excellent test to evaluate AI progress. Just being able to detect when a players is talking for him or for his character would be a prowess for an AI.

Will we reach the positronic brain of Asimov robots?
But then our concerns for the game aspect will be very small vs other impacts in the real life.


The Laughing One
When the AI throws an exception, it just defaults to: "Throw for Initiative!" (and in the background throws on the random encounter table)...

Thomas Shey

I've heard some reports of people using ChatGP to DM some modules, and the reports were that, at least in the context of running modules--they weren't bad.


CR 1/8
Here's an article about a research project in which several AIs are let loose in virtual world similar to The Sims. Not really a game at this point, but it's interesting to read how these GPT agents wander around their virtual world and interact, with themselves and with human users.

It's easy to extrapolate this to a DM's perfect little campaign world of NPCs before the PCs show up to wreak havoc. :devilish:

Related Articles

Remove ads

Remove ads