Will AI reach human-level performance in Magic: The Gathering before 2026?

1kṀ3384

2026

14%

chance

ALL

Resolves YES if an AI can beat me or a similarly-skilled player at least 4 times out of 10 games in a mirror-match. (The AI and I both have the same deck and the same amount of time, with at least 30 minutes per game.) Only the strategical decisions must be made by the AI; a human can do the card motions (if tabletop) or clicking and data entry (if online).

Any reliable evidence that this has occurred is sufficient: e.g. a paper published where the researchers claim to have run such a test.

The AI can't be designed for the specific deck it's playing, it needs to be general-purpose. It does get to know in advance of the game what cards are in the deck it's been given. Any reasonable deck for a serious format is fine; it can't just be Forests and Grizzly Bears.

Gaming

Competitive Gaming

Magic: The Gathering

Get

1,000

to start trading!

People are also trading

Will an AI by OpenAI beat a super grandmaster playing chess by 2028?

52% chance

Will AI beat top Magic the Gathering human player before the end of 2026?

11% chance

Will AI beat top Magic the Gathering human player before the end of 2028?

30% chance

Will AI be superhuman at MTG rules by the end of 2030?

79% chance

Will AI be passable at answering Magic: The Gathering rules questions before 2030?

88% chance

In 2028, will an AI be able to play randomly selected computer games at human level without getting to practice?

50% chance

Will an AI be able to play 3-person Monopoly Deal or an equivalent card game at a superhuman level by the end of 2025?

75% chance

Will Magic: The Gathering Arena introduce bots in multiplayer by the end of 2025?

30% chance

Will Wizards of the Coast change their stance on not allowing AI art before 2026?

44% chance

Will an AI system beat humans in the GAIA benchmark before the end of 2025?

Sort by:

In early 2024 I tried to play a game of MTG with chatgpt, it went horribly. it was so bad.

in late 2024 i tried again and it refused to even engage in the practice lol. It wouldn't play, said it couldn't or whatever.

these were all free versions mind you, and its an LLM, not even an ai trained in card games or magic or whatever. but it was sooooo bad this is like.. inconceivable to me

@No_uh Makes sense, see the terribleness of Claude plays Pokemon.

I suspect however that if an AI were to be trained via self-play on Magic specifically, it could resolve this market YES with current tech. AlphaStar was 6 years ago.

@IsaacKing I agree! Unless someonewith good seing sees this market, though... I'm on the NO still for 2025.

WHY NO FOREST AND GRIZZZLY? BEAR GOOD. FOREST GOOD.

The AI can't be designed for the specific deck it's playing, it needs to be general-purpose

Does this mean it needs to be able to play a deck without "practicing" with it before? If so, I expect this to be nearly impossible. Magic rules are incredibly compliated and the action space such an agent will need to learn without practicing is immense. On the other hand, an AI that can play a certain mirror but can easily be retrained to play other mirrors seems very plausible.

Some provoked questions, answer as you'd like: What is your level? And what might qualify for someone to play instead of you?
(How many times a year would you be up to take the challenge?)
How would the deck choice be made when running the test?

@ChameLeon Yeah, I feel like these are substantial ambiguities. I predict human level performance for an AI piloting mono-red before KCI.

@ChameLeon I'm the sort of person who consistently top 8s RCQs and never wins any.

No specific number.

Mono-red is fine, but it has to be a general AI, not designed specifically for mono-red.