Can LLMs play chess? Yes, but not in the way a chess engine does. A large language model such as GPT, Claude, Gemini or DeepSeek can name every opening, explain why a knight on the rim is dim and annotate a famous game move by move. Put the same model in front of a live board, though, and it can walk its king into a mate in one or try to move a bishop through its own pawn. This article explains why that happens, what LLMs are actually good at, and how you can watch them play full games against each other in the browser.
Why a language model can play chess at all
LLMs are trained on huge amounts of text, and a surprising amount of that text is chess: millions of games written in PGN notation, opening books, forum threads, puzzle collections and commentary. From that data a model learns strong statistical patterns. After 1. e4 e5 2. Nf3 it has seen 2... Nc6 countless times, so it plays it. In well-known openings a model can look like a strong club player for ten or fifteen moves.
What the model does not have is a board. A chess engine such as Stockfish keeps an exact internal representation of every square, generates every legal move and searches millions of positions per second. A language model predicts the next token. Any "board" it uses has to be reconstructed from the move list or from a text description, inside the same process that is also deciding what to write next.
Where LLM chess goes wrong
Once a game leaves familiar territory, the same weaknesses show up again and again:
- Illegal moves. The model moves a piece that was captured earlier, ignores a pin, castles through check or "forgets" that its own pawn blocks a diagonal.
- Lost board state. The longer the game, the more moves the model must track. Mistakes compound, and by move 40 its picture of the position may simply be wrong.
- Shallow tactics. Models often spot a one-move capture but miss the recapture, or play a "check" that loses a piece.
- Confident explanations. The written reasoning frequently sounds excellent even when the move is a blunder. Fluent prose is not the same as calculation.
Newer reasoning models, which think step by step before answering, do noticeably better at keeping track of the position. They are also much slower, and they still make mistakes that no engine would.
Can LLMs play chess better with the right information?
A lot of the illegal-move problem comes from asking the model to do two jobs at once: rebuild the board and choose a move. You can take the first job away without helping it with the second. In AI Game Arena's chess mode, every model receives the full, exact position as structured data, plus the list of legal moves and the direct, factual result of each one (for example, which piece a move captures or whether it gives check).
Crucially, it gets no evaluation, no score and no recommended move. The referee states facts; the model still has to decide which of twenty legal moves is actually good. That turns the test into what most people want to know: not "can this model remember the rules of chess?" but "can it think about a chess position?"
What still separates strong and weak models
Even with perfect information, models differ a lot. Some play purposeful, solid moves; others shuffle pieces or grab material without checking what comes back. Speed differs too: a fast model may answer in two seconds while a heavy reasoning model can think for over a minute on a single move. Watching both at once is a quick, intuitive way to see the trade-off between quality and latency.
LLMs vs. chess engines: a quick comparison
| Aspect | Chess engine | Large language model |
|---|---|---|
| Board representation | Exact, built in | Reconstructed from text |
| Legal-move generation | Perfect | Error-prone without help |
| Search | Millions of positions per second | A few lines of written reasoning |
| Explaining a move | Numbers only | Natural language, often persuasive |
| What it tells you | Chess strength | General reasoning under pressure |
That last row is why LLM chess is interesting even though engines are vastly stronger. Nobody needs a language model to beat Stockfish. Chess is useful because it is a clean, rule-bound environment where every decision can be checked, which makes it a good lens on how a general-purpose model reasons.
How to watch LLMs play chess against each other
- Open the AI Game Arena lobby and pick Chess.
- Choose who sits at White and Black. You can pick two different models, the same model twice, or take a seat yourself and play against an AI.
- Press Start. Each move shows the model's short comment, its full reasoning, how long it took, and any errors.
- Export the finished game as a PDF report, or record it as a video to share.
If a model returns an illegal move or its API times out, the game does not make a random move for it. The match pauses, the error appears in the log, and you decide when to retry. That keeps results honest: a model never wins or loses because of a move it did not choose.
Beyond chess: other games that test AI reasoning
Chess is a game of perfect information. To see how models handle hidden cards, bluffing and probability, try Texas Hold'em or the three-player card game Dou Dizhu. For pure pattern recognition, Gomoku and Othello are fast and brutally unforgiving. Want to know how we keep all of these comparisons fair? Read how we designed our LLM game benchmark, or browse the full list of AI players.
The short answer
LLMs can play chess, and the best of them can play reasonable, sometimes impressive games when they are given the position accurately. They are nowhere near engine strength, they still blunder, and their explanations are more convincing than their calculation. That gap between how well a model talks and how well it plays is exactly what makes watching them compete so revealing.