A challenge extended to AI deckbuilders
Commander (EDH) forum
Posted on May 9, 2025, 5:29 a.m. by legendofa
I like to think I'm reasonably creative with deck ideas, and that I can design something that nobody's ever seen before. I just spent some time checking out AI deckbuilders, and I was generally unimpressed. They do decently, if boringly, with well-established archetypes and common shells, but give them an unusual prompt and they start to fall apart.
I asked several different AI deckbuilders to design a Temur Goad deck for Commander. This combination has only four decks on TappedOut (and I'm happy to say one of them is mine), six on Archidekt, none on Moxfield (but that could be user error on the search), and 46 on EDHrec (assuming Forced Combat overlaps with Goad). So it's not exactly a common strategy.
I'm not going to reveal which AI builder gave which results, but I tested edhgen.ai, ai4chat, ChatGPT, and yeschat.ai a couple times each. The results included:
-
a Temur goodstuff deck with no goad cards.
-
a 64-card deck that included Taunt from the Rampart and Karazikar, the Eye Tyrant
, claiming they weren't legal as commanders but could be in the 99. -
Estrid, the Masked
as a highly recommended card, then quickly devolving into pure broken-AI gibberish. -
a suggestion for the best commander as Nefarox, Spymaster from the RRG set--anyone recognize that?
-
a Xyris, the Writhing Storm deck (yay!) that was mostly wheels and token stuff (boo!) that tried to sneak in Kardur, Doomscourge. (This one was the best, and I'll give it credit for recommending a card I had overlooked. It was still mostly generic first-draft-level Xyris with off-color-ness.)
-
an unfocused Marisi, Breaker of the Coil
deck. -
Raggadragga, Goreguts Boss + Grenzo, Havoc Raiser as a Commander + Background option, then claiming that green was off-color.
I then asked each one for a deck that uses only mythic rare cards and basic lands, a looser version of a prompt I got on here. Set symbols are red, but the budget isn't is where I ended up. Similar results for the AI:
-
nonexistent cards, then breaking.
-
a 38-card deck.
-
cards of all rarities, failing to understand the prompt.
So I'm offering a challenge to AI deck builders and those who use them: pick a prompt for a niche archetype or unusual deck gimmick and post it here. I build a deck meeting that prompt using my human creativity, the AI of your choice builds a deck using whatever digital skills it has, and the TappedOut community judges the decks.
Or, if you have other general thoughts or comments about AI deck building, please feel free to add to the conversation. Spoiler: I'm generally against AI-built decks. The best ones are generic and un-nuanced, the worst ones straight-up don't work as decks.
TypicalTimmy 100% agree on what you say for both the cause and the effects. The only point I want to mention is that while I was doing my thing, I didn't try to refine the decks at all--I took the first output they gave me, skimmed through it, then closed the session and started a new one. All the problems I listed were in the AI's first response. With that in mind, I'm sure that if I tried to ask the AI to improve the deck, it would just move farther away from what I was looking for. I might try that today and see what happens.
I should say I'm not entirely against AI. What I don't like is unreviewed, unsupervised AI being used in ways it isn't meant for. Generative AI is imitative and predictive; it says things that look right, not things that are right.
May 9, 2025 1:38 p.m.
TypicalTimmy says... #4
After my experiences, and with AI giving me entirely inaccurate answers to basic algebraic expressions, I have made a personal decision in my life that AI should be treated like Wikipedia.
It's a great starting point, but don't cite it and take it only at face value.
- Wait, how did AI mess up... math??
Rounding.
Let's say I have an expression and the answer for part of that expression was 34.344645921
How would you round this?
Sometimes AI was keeping as is. No rounding.
Sometimes the same AI input 34.34
Sometimes it rounded to 34.35
How? 34.344 rounds to 34.34? You round down because the hundredths place is also a 4?
Yes, but the thousandths place was a 6.
34.3446
This rounds to 34.345
Which AI then rounded to 34.35
May 9, 2025 5:39 p.m.
I'm a game dev, and I use gAIs a lot. Mostly for debugging ("Here is a thousand line call stack, find the error") or boilerplate ("Repeat the pattern public void SomeNameX() from 0-100).
Occasionally, I will ask it how to perform a specific action. "How do I authenticate a user in Unity C# using Google Auth?" kind of thing. It may or may not tell me to do some setup, but almost always it will give me a function that doesn't exist. When asked if that is real, it will say "Of course it is", when asked to provide a source it will go "Good catch! That is not a real method! Try this instead:".
If the gAI is capable of accessing the internet, I find asking it for sources is a good way to keep its response relevant. It doesn't always work, I've had it cite documentation from v1.0.1 of an API that is on version 13.2.3, and sometimes it gets fixated on a Reddit post that is just one person asking "How can I do this thing? I'm trying in this way" and will not consider other sources haha.
Not super relevant to a more open ended prompt, just wanted to vent!
May 10, 2025 3:02 p.m.
TypicalTimmy says... #6
This was the expression I wrote btw for a major project I have been working on. I can't get into details for it yet, but I finally figured out ChatGPT and other AI tools were rounding improperly or not rounding at all, so every single time I would enter the variables I would get radically different results.
Finally after 3 days I figured out the rounding error and corrected it.
My friend suggested I try to learn python for this, so that's in the works now during my very rare free time.
R(H) + F(R / 2) + X(6) + 4(R * S) + 140(B) = Y
Y / H = A
A(O) = L
L + Y = C
C - G = TOTAL
TypicalTimmy says... #2
I played around with ChatGPT recently to optimize theoretical builds in Dungeons & Dragons - purely for fun, as I walked away from the group I was in two weeks ago due to personal matters.
What AI will often times tell you sounds absolutely amazing, but it will get confused and give surface-level details.
I got into an argument with ChatGPT because it was contesting that a 13th level Cleric in 5e would have access to a single 7th level spell, but be wholly incapable of casting it. Yeah.
It was also misunderstanding rules for Action Surge.
It was giving incorrect damage averages.
It was misunderstanding spell rules.
What happens, that I've noticed, is AI will skim the Internet for information. Now, AI doesn't "read" so much as "predicts". This is a key distinction. It tries to assume what SHOULD be next based on what already was.
This is why AI devolves overtime. Because whenever it gets an input, it looks for past references for its output. The longer your chat history goes on, the more it looks back at its own communications in order to keep continuity.
But, it's a game of telephone. Eventually it'll make a mistake, then it'll reference that mistake, then it'll try to back-track on that mistake. But it's now using multiple mistakes to reference its mistake.
I'm more than certain if you ask it to build a 100 card deck, it'll do it just fine. But as you ask it to tweak and hone the deck, it'll begin to make illegal decisions, misclassify cards, not understand the complex rules systems, give you faulty information, etc.
One of the things it did for me was attest that I could do an average damage output of 177 per turn indefinitely.
Well, no. That's only happening because of burning spells thanks to Divine Smite. Once those spells are burned, that damage ends.
It said I could get 5 attacks per turn through the attack action, action surge, extra attack and haste.
Well, no. You can't action surge during your attack action, and then your extra attack, and then action surge the 2nd turn again and then the 3rd and the 4th... You have to finish a short or long rest.
Meaning action surge is a 1 time use.
See, AI has the unfortunate reality of accumulative information. If there is 1 post (the DMG or PHB) that says how Action Surge works, but 100 Reddit posts on how to "use it" and they are all wrong, it's going to see 101 documents of Action Surge and base it's opinion on the 100 incorrect posts, and not the 1 correct one.
So, just bare that all in mind when attempting to use AI to build a deck, ESPECIALLY if it is one you intend to then buy.
May 9, 2025 8:15 a.m. Edited.