Will AI reach human-level performance in Magic: The Gathering before 2024?
Standard
114
Ṁ24k
resolved Jan 1
Resolved
NO

This market resolves positively if either of the following is true:

1. Someone publishes a paper where AI estimated ELO in at least top 10%, with player pool of at least 50 players. The games should be played by rules one of the officially recognized MTG formats and players should not be restricted in their deck choice any further than the format requires.

2. AI gets at least third place in any official MTG tournament with at least 50 players.

3. AI reaches the top 100 rank in MTG Arena.


Edit: Human-designed decks for AI are allowed for the purpose of this question.

Get
Ṁ1,000
and
S1.00
Sort by:

@MikhailDoroshenko Can you resolve?

predicted NO

I'm not sure if the present version of the Comprehensive Rules allows for AI play in official tournaments because it would probably count as Cheating to have a human interface with it

So people would be basically breaking either the ToS of an online client or the regular tournament rules of a paper version by interfacing with AI's - you'd want to have a chat with senior judges before doing this anywhere than the kitchen table or a deliberately casual tournament.

predicted NO

@PipFoweraker That would not be allowed in a sanctioned tournament, no. The rules define a player as a "person", so until an AI is legally recognized as a person that's not gonna work.

That doesn't stop people from running unsanctioned tournaments where AIs are allowed.

predicted YES

@PipFoweraker #3 - AI reaches the top 100 rank in MTG Arena is plausible from an AI that operates without official blessing simply by grinding at a slightly positive winrate. Whether or not such a candidate already exists (you do not need or want top 100 for most purposes) is an interesting, if difficult to answer, question.

predicted NO

@TravisLuckenbaugh You can get pretty far on MTGA just by playing a lot of games, but idk about top 100

predicted NO

@Julian I think he meant "reach mythic rank" reaching top 100 just by having a winrate slightly above 50% is not really possible

predicted NO

@Odoacre I mean, you can if you maintain a positive winrate against actual top players, but that's a much more difficult ask.

predicted NO

@Julian I guess you are right, that's the very definition of being a top player. I thought by "grinding" you just meant playing random people online.

I'm holding out for the AI that can outplay Menendian in a Vintage Gush mirror

predicted NO

Another market that gives it an extra year:

Capabilities won't be there any time soon. All complex-game AI to date have had serious problems dealing with complex hidden information, where what actions the opponent can take depends on the hidden information. Magic is basically the most complicated game possible on this front.

Whether or not the capabilities will be there in the next year, my guess is that no one who could build a good enough AI will bother to actually do it.

I'm not sure those criteria actually require human-level performance. Some of the formats allow for successful decks with minimal interaction - which means you simply have to design an AI that can pilot a strong archetype well enough. You could probably give it some very simple heuristics based on mana value or total number of triggers to improve its ability to wield removal, too.

I'll admit I'm not sure exactly what criteria would avoid this. But I suspect that you have to compel the AI to also build the deck if you want to force it to show a "true" understanding of the game.

predicted YES

@AndrewHartman True, but people in the tournaments can also use a decklist they did not invent themselves. The purpose of this question is exactly "how hard it is to build a bot that would pilot some archetype really well against a diverse set of opponents".

@MikhailDoroshenko Well, depending on the block, the pool of opponents might not be particularly diverse, y'know?

predicted YES

If anyone is not ok with my edits for this question, you can ask me for compensation. I still want this question to be open, so I changed the resolution to be more in line with the result I wanted to test.

Criterion #3 has already happened for some definition of "AI". I remember a reddit thread where someone got to mythic with a human-designed deck played using a deterministic algorithm - always play the leftmost card in your hand first if possible, always attack.

This happened because if you keep losing on Arena, the system matches you with worse and worse opponents until just having a reasonable aggressive deck is enough to win.

predicted YES

@Multicore Big oof. Will Mystic top 100 be better or is it still vulnerable to this?

@MikhailDoroshenko Top 100 mythic would be a lot better, yes. Maybe with a restriction of not being in the first couple days of the season to avoid the case where the bot rushes to be one of the first 100 while abusing an already low matchmaking rating.

There's an extent to which maintaining a high mythic rating is more about grinding out lots of games than about being maximally skillful, but the bot would still have to have a winning record against high mythic players.

predicted YES

Does the AI have to do all parts of the game? For example, what if there is a computer that decides on all the moves, but uses a human-designed deck?

2nd question: anyone knows whether MTG arena allows bot players, or would that violate its terms of service?

@Sjlver Good questions!
For the first one, I would like to allow for a human-designed deck, because it is a drastically different part of the game and the research paper that tackles one aspect might not pay attention to the other. (I will specify it in the question)
For the second one, I tried to search for this but did not find any conclusive results.

predicted YES

@Sjlver One more reason to allow for a human-designed deck is that the opposite to this question (human play with ai deck) has already happened, so it is not a very interesting restriction. https://www.reddit.com/r/MagicArena/comments/s1i0ud/i_made_it_to_mythic_rank_32_using_a_bot_to_draft/

predicted YES

@MikhailDoroshenko There are some relevant (but also quite vague) bits at https://company.wizards.com/en/legal/terms

2.2. Restrictions. You agree not to engage in any of the following with respect to our Websites, Games, or Services:

(i) Data mining: Use any unauthorized means, process, or software that accesses, collects, reads, intercepts, monitors, data scrapes including without limitation, agents, robots, scripts, or spiders; or mines information (including reverse look-up or attempted tracing of Registration Data in any way and for any reason);

(v) Cheating: Creating, offering, using, promoting, distributing, or making available any cheats, bots, hacks, or device, software, code, or program not expressly authorized by Wizards that grants any user an advantage over other players not using such methods;

Anyway... not sure these would really matter; I presume researchers would reach out to Wizards and ask for permission.