I wrote about this in late 2024 in a piece that took me a while to get together. Now, that took me quite a long time, because I wasn’t sure where I would like to advocate those tools.

This is more about people are going to use them anyway, so why not try to push them into using them beneficially. I’m not going to contradict myself from the previous post, this is more like a more thoughtful take on the topic.

We Don’t GM Despite It Being Difficult, We GM Because It Is Difficult

According to certain fields of philosophy, in life, there are instrumental goals and intrinsic goals. Instrumental goals are the things you need to do to be able to do the intrinsic goals. So, intrinsic goals are the thing you want to be doing, while instrumental goals are the means to an end. So, you eat and you sleep and you work as instrumental goals to be able to do the fun things.

Now, why would you now avoid the fun thing by having an AI do it? For me, so much of the GMing experience is about thinking about all the possible scenarios you could bring your game to. Another huge chunk is the research. Just figuring out how things work in this world specifically or in our world so that you can make your own version of it. Then there’s creating the characters, scheming on their behalf, thinking about the music, and so on and so on. I think I’m on the lower end of how much time people use for preparation, because I mostly just give my players choices on what to pursue, but I still use quite a bit.

So, using AIs to do the work for you is just going to take away what makes you happy or content in life.

If You Are Lazy Enough to Let AI Do Your Work, You Can’t Be Trusted To Not Be Too Lazy To Check the Work

And this is a big deal. So much of what we, as a species, have evolved into is going down the toilet because we are using AIs. I mean, there’s a study which found that the two most commonly linked words to colleagues who used LLMs were ‘lazy’ and ‘replaceable’. Do you want to be that to your friends?

Sure, is there a difference between an LLM and a scenario written by someone else? Actually, yes. Supposing that scenario was written by someone qualified, it will have a flow and ideas that the LLM can’t reproduce. However, the bigger difference often is that people don’t actually read what the LLM says. Those are so dry and boring that people don’t take what they say in. They just skim through. So, basically they just trust it to be right, even though they know that it constantly gets things wrong. That’s not a good place to be.

On a Technical Note, There Is Such a Thing As Context Windows

What is that? A context window is basically how much an LLM can maintain in it’s memory at the same time. I don’t know what it is exactly for each LLM these days, but it can fluctuate based on load. What this means in practice is that when the LLM hits the ceiling in one conversation, it will start to summarize things but we don’t know what is being cut off. So, suddenly important things are forgotten or unimportant things become very important because now they occupy a larger part of the memory.

Again, I don’t know exactly how much of this memory there is in various LLMs, but at one point OpenAI was marketing 40k tokens for paid version and 20k tokens for the unpaid version. What does that mean in practice? Most tokens are (these days) singular words and a typical novel is somewhere between 60k and 90k words. So, an LLM can’t hold a single book in it’s memory. How are you going to explain a whole world to it? How are you going to make sure that the whole history of your campaign is there, when there might more information than the context window limit or maybe the limit was cut to 5k for a moment when you happened to use the LLM at a peak hour and a bunch of stuff is lost.

Sure, it will remember all that stuff later on, but it has worked based on the lower limit already, which will have affected it’s thinking in number of ways.

LLMs Have a Hard Time Knowing What’s Important

I have been writing a book for a while. Part of the reason is that I like writing (those intrinsic goals again) and I need a project where I can stress test LLMs. So, I was testing the LLM used by Adobe on Acrobat. I can’t anymore, because the usage limits are quite low, but I gave the LLM my 80k+ book and asked it to list the 10 most important characters, just to see whether it understood what was going on. (Side note: I tried a similar thing with ChatGPT and it didn’t even realize that Kimberley and Kim were the same person.) The Adobe LLM didn’t really get everything, but it got a decent list out. However, it suggests further topics of discussions, as LLMs often do. One of those questions were about the importance of the puppy in the book.

Sure, there is a puppy in the book, but the Puppy comes in at the very last minute. It’s there basically for a joke on gender expectations (a kid says cats are girls and dogs are boys), but somehow the LLM managed to feel it was very, very important. This might have something to do with the aforementioned context windows, as maybe it had purged everything up to that point already and it didn’t remember much of the book, but it still felt weird that it raised that specific topic as something important to dwell into, especially considering that I was using the LLM for a purpose it wasn’t built for (it was built for summarizing text rather than editing it, which was what I was doing, basically).

This Doesn’t (Necessarily) Mean You Shouldn’t Use Them

Although, the moment you feel you can’t do something without them, you need to stop using them right there. I am being serious. That is a way to protect your brain. There is a lot of evidence that using LLMs can cause you to lose your ability to think critically. And that’s a pretty important skill.

LLMs are good for ideation, but you need to be clear with them that you want it to be critical and this can be difficult. They often default to being nice (which is probably a large part why they kill the critical thinking part of your brain). So, in order to get the machine to be critical of you, you have to push it in that direction. So, give it ideas and ask it to make a SWOT of it. This might not work with everything that well, but it will try to find weaknesses and threats in your ideas. You can also ask for one-star reviews of your work or ten ways to do it better. Just don’t let yourself fall into the trap of feeling good because a machine designed to make you feel good is trying to do that at the cost of your ability to think.

Finally

I will leave you with this minidocumentary on Lightning Bolt, one of the best noise bands around.

So, go out and do shit because you enjoy doing shit. You owe it to yourself to do better than rely on some machine that is trying to steal your enjoyment of life.