AIs in Aid of GMing

Can AIs be helpful for GMs? Yes, but not for very many things.

Before we go deeper into this, what you need to understand is that AI can never understand the actual experience of being a human. It can only learn about what it can digest as data from some source. Since it doesn’t have access to very basic things about the human experience, since we don’t feel the need to document it, it can never learn about that. Example: AI finds the data that something specific tastes like chicken. So, now it knows that something tastes like chicken, but it has no idea, what chicken tastes like. The AI can never understand what it feels like to walk or to touch grass (not figuratively). So there is always going to be a limitation on what an AI can do for you based on this gap.

But also, AI is not creative. This is somewhat difficult. Creativity, in general, is about taking existing concepts and combining them in new ways. Can an AI do this? Yes, but that is not the entirety of what creativity is. This new thing also has to fit the purpose in order to be really creative and this is another thing AI fails at, because it doesn’t understand how things fit together unless it has been able to learn it from somewhere, so it can’t verify how well something would work. So, if you want AI to help you with creativity, you should always as for a list of ideas and use those.

Finally for these caveats, be very careful about publishing anything with AI input. Sure, you can use AI to help you come up with ideas, but it is always stealing those ideas from somewhere. It just won’t tell you where, so you have to check yourself, so that you are not plagiarizing by accident.

With these limitations in mind…

I have never had as much respect to D&D players as when they stood up to Wizards regarding the leaked predatory changes to Wizards’ services. I am not a D&D player myself, so I might not be fully aware of all the things that went on here, but one of the things that did not receive very much attention was Wizards’ proposal to let higher tier customers use AI to run their games. That sounds horrible to me.

First, that means that whoever was behind that decision does not have any understanding of how RPGs work. Sure, the idea was that the AI would only run combat, but as I understood this, the whole thing would be just a series of combat encounters, because the AI can’t really handle anything else. Perhaps there could have been something else in between, but it could not have been very complicated.

So, how would an AI do in this situation? Let’s take the ChatGPT approach:

ChatGPT does not understand rules and neither does any other AI using similar methods. They are for language modelling, so they will regurgitate language without understanding the meanings behind those words. So, it’s actions would be based on what has happened in earlier games and it would mimic things, but would not understand the context, so it would be completely unpredictable and unable to understand the environment, as well as other limitations it should be aware of. So, you can’t do this.

But, you might be thinking, computers are good at playing games. Why couldn’t you just use an AI that’s good at that instead of ChatGPT? Yes. This is true. You can teach an AI to understand combat to such a level that it could do this, but it’s not easy. Take Chess again. That’s a fairly simple game. There’s only six different pieces that each has it’s own pattern. Despite this, the AIs used to play chess often rely on identifying the game state and choosing a move based on that. That requires playing endless test games. You can’t really do that with RPGs. The maps are more complicated, the number of possible moves is exponentially higher and there is randomness, which means the learning process needs more time.

Sure, you can still use an AI to do this, but it has to have a very different approach. There is a way of approaching this, where the AI tries to form a view of the world, sets goals based on what it sees and then makes plans based on those goals. This makes the whole thing somewhat easier, if the AI can easily cut the amount of possible actions down from thousands to just a few. It will also make the combat feel realistic, as the AI can make plans over a few rounds of combat instead of just one at a time. However, players are complicated. Forming this kind of a worldview requires the AI to have an idea about what the PCs might do… and that is hard.

And with all of this, you need to let the computer learn. What is it supposed to learn? The thing about DMing is that you are not supposed to make things challenging. You are supposed to make things feel challenging. That’s the sweet spot. So, you can’t really teach the AI to try to win, but the AI can’t know when people feel challanged, so how is it going to calibrate it’s actions?

In the end, instead of using AI, you would just go back to the simplistic systems that have been used in various action based games since forever, where each NPC has some simple rules it follows, which just makes the whole thing boring and predictable.

So, we don’t want that. What else could an AI do for us?

Some bigger video game studios have started testing AI-NPCs in their worlds, which would be able to speak to the PCs using AI generated lines. I am not hopeful about this. There needs to be so much information regarding the world for this to work that it is practically impossible. At the same time, not every NPC can know the same things. Most people don’t know very much about the minutiae of what’s going on in the world, so in order for this to work, each NPC would need to have their own models and they would have to be interconnected in a logical way. There would need to be some sort of simulation of how information moves in this network as well, and the emergent system this would provide, would need some way of making new NPCs, who have to feel like they belong as well. You just can’t make these characters realistic. Sure, they can appear outwardly as such, but trying to maintain any sort of discussion with them will just fall apart.

What ChatGPT can do for you is be a quick way to figure certain random things out. Sure, you can ask it for things like scenario ideas, but those will be unimaginative. They are going to be pretty much based on generic ideas from existing sources, so you can do that easily on your own. In general, if I ask ChatGPT for these kinds of ideas, they are often just blatant plagiarism, which is fine if you are not planning to publish anything, but since these are often so blatant (once ChatGPT named the villain Doctor Doom, in another case, a setting reminiscent of Edwardian London was just London in 1910, which is the end of the Edwardian Era), you will have to just change things before you can use them and since they are so uninspired to begin with, this is just not a good way to work.

ChatGPT does not plan what it writes. When a human writes a sentence or a paragraph or anything longer, we have an idea where we want to go, at least to a point. ChatGPT does not work like that. It’s based on analysing billions of pieces of text. It splits those texts into four letter tokens and makes statistical predictions on what four letter token would follow the previous token in a certain context. So, it doesn’t actually understand what’s going on. The system just tries to recognize the context based on a few keywords and after that it’s just basically math.

But if you need to a have list of short character descriptions for random people the players might meet in a specific setting, ChatGPT can produce that. You might just have to milk it a little, if your setting is not easily explainable to the AI.

The best way to use AI in this context is images. The setting might again be a limitation, but as someone, who can’t really draw, this is an easy and quick way to level up your game. Sure, you can also google for images, but there are going to be limitations as well. AI tools enable you to also make images with techniques appropriate for the situation. Think cave paintings, wood carvings, stained-glass, or even photographs with weird elements.

Admittedly, I am not that familiar with these tools. The one I have been using is called Freepik. Like all the ones I’ve seen, it has a limited number of images you can create and it also seems to have a problem with server capacity, as so many requests don’t seem to go through. I don’t know if it is any better than it’s competitors, but it just happens to be the one I stumbled upon, when I was looking for one. I just didn’t realize how limited the number of images I could create was, so I kind of wasted those usages on just testing how well the system works and how well it understands what I want. As there are plenty of these tools and they all work in much the same way, you can just find new one, when you need new images.

The problem with limited number of images is that what you often find yourself doing is tweaking the images. These don’t count as full images, but these services keep a tab on the amount of server time you have used for your requests, so each edit will eat some of this but not necessarily a whole images worth.

Of course, if you have access to premium on one of these services, use that, but those can be quite expensive, especially if this is the only use you have for them.

There already exists some websites, that try to employ AI to produce something, but they seem to be more or less based on ChatGPT, so you might as well use ChatGPT by itself rather than add the extra level between you and it. Those sites seem to often have problems anyhow.

Final words:

AI can be helpful here, but it can’t really offer the experience you want from an RPG. They can be used for a variety of things, but basically for anything text based you will probably find another usable tool on the net anyhow. Where AI can shine is tha images. Just a few good pieces that give your players a good idea on what something in your world looks like can go quite far.

Leave a Reply

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

This site uses Akismet to reduce spam. Learn how your comment data is processed.