An offshoot from another thread where there some discussion of ai’s playing chess.
Obviously, ai’s can play chess and other board games. And they can beat humans.
My question is whether the ai’s have to be programmed to play these games. Or have ai’s reached the point where they can read the rulebook to a game they have never encountered before and develop their own strategies of how to play the game?
I’m talking normal games not simple games that can learned with beads and matchboxes.
That’s basically describing the best Chess AI engines.
About 10 years ago now, AlphaZero, a neural network based chess engine appeared. It had been trained basically by playing itself. After playing millions of games, it figured out what was good play vs bad play. And it was significantly better than the previous generation of engines (including the successors of Deep Blue) that were more strictly algorithmic.
That’s probably slightly different from what you are envisioning. That’s not an a priori strategy based only on the rules but one developed mainly from actual game experience.
Many of the advanced chess engines now do a bit of a mix. There’s still tree searches and so on but most of the top engines incorporate at least some AI elements.
ETA: An article from back in 2018 about it:
Yes but I’m assuming a human had to “teach” AlphaZero the rules of chess first.
My question is whether Ai’s have reached the point where they can teach themselves how to play a game without needing a human.
Well, sure, but just the basic rules of the game. You have to teach a human the same. A human might be able to figure out most of the rules from observation but we’re very likely to miss esoteric ones, like taking en passant or castling rules or whatever.
But the answer is more or less ‘yes’. Or at least there’s a fair amount of overlap between a not very bright human and the best neural network algorithm at being able to figure such things out. Might make mistakes but so would many people.
I’ve learned how to play games without being taught them by another person. I learn by reading the rulebook.
Can an ai do this?
I’m not sure I really understand the question.
Reading the rulebook is equivalent to being taught by another person, just not with the person in front of you. It is functionally equivalent to being programmed the rules by other means. The “how” is different but you are still being programmed in chess by another person. The difference is you chose when, where, and if to learn the rules.
So if the question is if “AI” (which is a bit of a nebulous term) has free will to do the same, the answer is no, that’s not possible. It can’t choose when, where, and if to learn the rules.
I don’t see why not. This is easily testable yourself if you really want to know. Pick a boardgame that you have, upload a pdf of the rules to your favorite Ai program, and then play it. Depending on the game you may have to describe or take/upload photos of the game progress so it can make decisions on its next move(s) but the process is pretty straightforward.
I don’t see how this works.
Take a modification - invent your own boardgame and make such a pdf of the rules. It’s not going to know how to play. It’ll just spit your words back at you or borrow from a different game that was part of the material it was trained on.
With existing board games, there’s a whole internet worth of information out there for it to work with and make it sound like it ‘knows’ what is going on. Though, this might actually be what the OP was looking for. If AI’s can actually parse and understand content. And the answer is, not really, but they’re really good at making it sound like they do. Which again, isn’t far from what a lot of people already do.
Well, ChatGPT has played chess. It is not good. In fact, it tends to make a lot of illegal moves.
It may have gotten better lately though, but probably not.
Here is a recent article about it: https://time.com/article/2026/04/13/why-i-play-chess-against-chatgpt/
It sounds like you are still under the impression that Ai are just fancy text predictors. A game with well-defined rules is more than enough for it to to make decisions on moves based on those rules and the information provided it. it may not play it well or with the best strategy, but it should at least be able to complete a game.
As the post above yours noted, ChatGPT doesn’t really do that, either. It can’t currently be relied upon to even complete a (legal) game. It can certainly sound like it does, though, to somebody who doesn’t know the rules.
AlphaZero is certainly more than a text predictor but also, it can’t read the rules to a new game and play, either, the way a ChatGPT can fake it. Different AIs (again, bit of a nebulous catch-all term for different things), different results.
Sorry if I being unclear. My question is what are ai’s capable of. Can an ai read a rulebook and then figure out how to play the game within the rules of the game? Is an ai capable of following a set of rules it has read? Is an ai capable of developing a strategy of what it should be doing in a game to win once it knows the rules?
It really depends on what model you are using. Like I said, ChatGPT tends to cheat at chess - and not always even in it’s favor. It isn’t really built to have a logical model that is game-rules specific.
You need a model that is built differently - around the rules of the game rather than the rules of human speech.
You can easily test this out. Tell ChatGPT a chess position and ask it what move it would make. It is very rarely the best move as determined by an actual chess engine.
That said, thinks like AlphaGo and AlphaZero show that a neural-net framework that is only fed the rules of the game and then just learns how to play organically can easily beat human players at any game.
The best way to use a general-purpose AI to play chess would be to tell it to use Stockfish or whatever and then play a game against it.
I’d be really interested to see the results. Choose a semi-complex game like Catan, or Agricola. Has anyone actually tried this experiment? I’d be moderately surprised if the AI could accurately interpret photographs of the board, and very surprised if the AI could play effectively, just from reading the PDF of the rules.
Can an AI learn to play a new board game based off the rules? Almost certainly. Can a consumer LLM like ChatGPT or Gemini or Grok do it? Probably not. Part of the issue is that, the longer the game goes on, the more the AI will forget the rules due to discarding old tokens in memory. Likewise, it will forget earlier moves or start conflating stuff. This is less a limitation of the AI in a general sense and more a tech limit due to hardware allocation.
This doesn’t mean it will play well or optimally. Obviously an AI trained on a bunch of games will do better than a general AI working off the rules. But then a master chess player will beat a dude with his first CVS-purchased chess set as well.
Chess is not a simple board game, so I don’t think that’s a fair comparison. There’s a lot more long-term strategy involved. Even so, in the article, it was mentioned that it played along decently enough at first before hallucinations. A board game that is more linear and doesn’t have near limitless possible strategy-based iterations would be much “easier” for it to play through. Roll the dice, make a move with your piece, and make 1 or 2 decisions per turn.
I’m intrigued enough that I’m willing to test this out with one of the games I have and see how it goes. If I’m wrong, I’ll freely admit it.
What AI specifically are you asking about? There is no single “AI” that we can answer this question about.
Generally I think the answer is “no” for things like ChatGPT and other LLMs. It doesn’t really have the logical framework in place to turn the text of the rules into a set of possible legal moves. It will almost always make mistakes, cheat, or otherwise violate the rules.
But could you create an LLM-based agent specifically to play Catan? I would imagine yes with enough training. It would just need a much deeper history and be trained rigidly with what the rules of the game are. Then you would just have it play itself over and over until it figured out what the optimum moves are.
Honestly for almost all games the level of complexity is so low that a specific-built program for just that ruleset is more than enough to beat any human player. It’s only for very “deep” strategy games like Chess and Go that this has been a hard problem for computers until the last few decades.
So as a quick test ChatGPT certainly thinks it knows how to play Catan, and it regurgitated the rules pretty accurately. It even offered to play a game and simulate the opponents.
If I have some time later maybe I’ll give it a go - not sure how the board layout stuff would go. I’d probably need to get out my version of the game and play along.
I just started a game of Yatzee against Claude.
You should probably use Claude for this kind of thing.
I would imagine it’s really good at Yahtzee! There is, I believe, an optimal strategy for that game that is probably available in it’s training material.