Magic: the Gathering

Opinion

ChatGPT AI to play MTG better: How it works, pros and cons

, Comment regular icon0 comments

If you're a technology fan, you might be aware of the new artificial intelligence that's revolutionizing the search tool market. But how far can it go when it comes to helping us with Magic: The Gathering?

Writer image

translated by Romeu

Writer image

revised by Joey Sticks

Edit Article

Have you heard of ChatGPT? Maybe not, but you've probably heard of OpenAI. OpenAI's initiative is to allow anyone to use artificial intelligence to contribute to their daily tasks. The process is generally simple for the user: you send a message, and it responds with what you asked for. It was released almost a month ago, so everything is quite new.

OpenAi has made several tools and modules, and one of them has become famous recently, which is the AI that generates images. This one has become controversial in the artistic world and is even bringing serious discussions about whether the machine is stealing human work or actually creating something new. And yes, most applications that generate your image as anime or Van Gogh are actually nothing more than applications using the OpenAI API. Oddly enough, this new reality can't be transplanted into Magic, or can it?

Ad

Come on ChatGPT, introduce yourself:

“Hello, I am a language model designed to assist users in various tasks. I have no personal experiences or feelings, but I am able to generate human text and respond to user input. My knowledge is based on a wide variety of texts up to 2021. Can I help with anything else?”

- ChatGPT's response

If you want to test the ChatGPTlink outside website on your own, just click here. But the purpose of this article is:

- guide you on how to use this new technology in Magic

- figure out what you can ask that would be interesting

- figure out what not to ask because it will generate machine error

How ChatGPT works

ChatGPT works similar to a robot assistant, in which you ask for an action, and it returns a response. The simplest example we can give is to ask the robot something.

Image content of the Website

Above, I asked who won the Magic: the Gathering World Tournament in 2018. It correctly replied that it was Javier Dominguez, but also went much further in the answer, providing information on how the World Championship works, which company owns the rights of the game and much more.

ChatGPT has been trained in so much that it understands the language of the Magic: the Gathering game, that is, the words we use the most.

Image content of the Website

The way it answers the question "What is an Esper deck?" exploded my head. First, I didn't need to say that it was about Magic: the Gathering, that was already implied in the previous question, which demonstrates that it is indeed a conversation. Second, its answer was from someone who understands the game, using language like "Control Deck" and explaining what an Esper deck actually is in Magic, again correctly. Third, I wrote it wrong, my English is not the best, and I wrote "a esper" and not "an esper"; and even so the robot managed to translate my intention correctly.

From this answer, I was encouraged to write this article. What else could it do for us?

ChatGPT caveats and limitations

First, let's just leave a few caveats. This tool is in Beta, that is, it will still be improved. You can see next to each robot's response there's a thumbs up and a thumbs down, that is, we ourselves are helping to make the robot "think" better.

Also, the program is currently limited to English.

Now look at the response below:

Image content of the Website

I ask what is called a deck with Island, Swamp and Plains and I get the answer: it's a 3-color deck. Okay, kind of obvious, right? Sometimes the tool's answers aren't perfect, so we have to understand its limitations and how to ask better questions.

We should also remember that it was trained with a data set until 2021, meaning it won't be able to help us with the current Magic Metagame. That might be why it focused on saying that Esper is a control deck and didn't mention that it could be a human aggro deck, for example. The trend is for the training set to become more current as they improve ChatGPT.

Ad

As a result, I decided at this time to focus on questions focused on the Pauper or Modern format, since some decks of these formats are still present today.

Questions about the Metagame

If you're completely lost in the game, you might end up asking that basic question: "which format should I play?"

Image content of the Website

ChatGPT is clear about the main characteristics of each format, and I believe it is beneficial for a completely lost person. But it is noticeable that it doesn't mention all formats, perhaps only the most "popular". But where is Pauper? Let's test it by asking about the 10 biggest staples of the format:

Image content of the Website

Yep, the answer was correct for 2021. It's amazing that it understands the term "staples" and even the term "Pauper", and I haven't used any of these so far in the conversation. However, it points out Daze as a staple, which was banned in 2019.

The unanswered question remains: which is the best deck for Pauper?

Image content of the Website

The Mono-Black Control answer was surprising. I wondered, could it be that there were still people playing with this deck in 2021? I asked Romeu, one of our writers, and he mentioned Mono Black had a sort of comeback when Chatterstorm was taking over the format because it had a good disruption package.

Questions about your deck

A very honest question that can be asked is: give me an X decklist. Below, I ask for an Izzet Phoenix without informing the format.

Image content of the Website

It answers us with a decklist that has two problems: it has only 55 maindeck cards, and it has Faithless Looting, banned in the format in 2019, despite it not knowing whether it was banned or not at the time the AI training was done. See the decklist below:

Loading icon

Well, I wanted a list of Pioneer and not Modern, so I decided to be a bit annoying with the machine.

Image content of the Website

Unfortunately, it gave me a Modern list. So, here I started to see other limitations: it doesn't identify formats that well. And I continued pointing out its mistakes to see if it ever got it right:

Image content of the Website

Unfortunately not. The bot wasn't able to give me a Pioneer decklist. And every time I pointed out illegal cards in Pioneer, what it did was remove the card from the deck and exchange it for one from the sideboard, making the deck smaller.

At this time, I decided to go with Jund, things never go wrong with Jund, right?

Image content of the Website

It managed to send me a Jund decklist, but again only 55 maindeck cards, so I don't know whether it has a problem with deck size or if it's giving me some flex slots.

Let's go to the next topic then. I decided to ask for another different Jund decklist:

Ad

Image content of the Website

The bot was able to provide me with a similar Jund decklist, only this time the focus was on Scavenging Ooze and didn't have Bloodbraid Elf.

Advanced and interesting questions

From then on, I decided to do more interesting tests. What does every Magic player want? Someone giving advice on their deck and helping them on how to play it.

So, I put my list in TXT format and asked, "do you have any suggestions for improving my deck?"

Image content of the Website

ChatGPT was wonderful on this answer. It broke down each type of card, giving suggestions for staples for each one. I put a Jund list without Assassin's Trophy, and it suggested adding it in case I want more removals. It was also interesting to see the additions it suggested in case I want more card advantage.

You can add a decklist as well and ask what the purpose of the deck is. I added an Infect decklist:

Image content of the Website

ChatGPT acknowledged that it is a Simic Infect and the deck's main strategy.

Something I noticed is that it doesn't identify a deck with 2-3 cards. But if you provide 4 or more, it can tell what the deck intends to do. With three or fewer cards, it tries to guess the deck like BG-X or Abzan and already informs that the deck is probably playing to generate value:

Image content of the Website

In addition to all this, you can order Sideboard Guides from the tool.

Image content of the Website

It will tell you how to play against each deck. I recommend in this case to use the word "Extend" or "Expand" that will make you understand that you don't want a general view, but a guide against each deck.

Image content of the Website

Can you see any errors in the guide above? It constantly mentions that Thoughtseize is good to sideboard against many decks like Affinity, and spoiler alert: it's not!!

Be cautious with these questions

I realized the deeper we dig, the more mistakes appear. Be careful with how-to-play or sideboard guide requests.

I, as I hate Infect, asked questions about how to play against this deck. See the answers:

Image content of the Website

Image content of the Website

To say that Grafdigger's Cage is a good sideboard against Infect or even that Tarmogoyf and Gilded Lotus generates an Infect immunity effect shows how this machine doesn't, in fact, play Magic.

For very complex questions, it might be best to still rely on Magic professionals and content creators. ChatGPT reminded me of the following meme with these answers:

Is ChatGPT an Adventure Time fan?
Is ChatGPT an Adventure Time fan?

Conclusions about the new tool

It got me hyped. I can't wait for this tool to leave Beta and start being trained with a real set of data, until the end of 2022. I imagined having an assistant that helps us by knowing the deck our opponent is playing with only 3-4 cards. There is a lot of potential for this tool as its understanding of Magic's core concepts gets further expanded.

Ad

Until then, be cautious. The tool is great to get some ideas, something to start with, or even something to improve on your deck. From its ideas, talk to your group, see if there's anything absurd, in the end, it's still a tool that can make mistakes and indicate you a few banned cards!