Can ChatGPT play chess?

Can ChatGPT play a full game of chess yet?

Sometimes, and still not reliably. The best public evidence is Game Arena, a benchmark Google DeepMind and Kaggle run and described in a paper submitted in September 2026. Its harness gives every model the current position and the whole move list on every turn, rejects any illegal move, and allows three retries before the game is scored as a loss. That is far kinder than a chat, where nobody re-sends the board and nobody rejects a bad move.

Even so, the models fell into clear tiers. Gemini 3 Pro and Gemini 3 Flash led on the arena's own rating scale (1325 and 1297), OpenAI's o3 and GPT-5.2 came next (1009 and 933), GPT-5 mini sat at 525, and Anthropic's Claude 4.5 models came last, Opus at 236. Those numbers measure models against each other, with the weakest pinned at 0; they are not a FIDE or chess.com rating. ChatGPT's current default may be newer than the models tested.

The errors haven't gone away either. GPT-5.2 still needed occasional retries for illegal or malformed moves, and the weakest models needed more than five a game, often more than ten. The paper finds they cluster after the opening, once the position stops resembling memorized games, and names a common one: the model is in check and proposes a move that doesn't get its king out of it. In the opening the models play plausible book moves, because millions of annotated games taught them what a normal first ten moves look like; the further a game gets from memorized theory, the faster it comes apart.

Why the numbers punish it: a chess game has on the order of 10⁴³ legal positions and roughly 10¹²⁰ possible games. A real engine navigates that space with an explicit board and search. A language model stores none of it — it re-derives the position from the written move list every single turn, so one misread compounds into a game that has silently diverged from the real one.

Why does ChatGPT make illegal chess moves?

It makes them for the same reason it loses the board: a language model predicts the next token from prior text, with no persistent representation of the sixty-four squares. Each move is re-inferred from the notation written so far, so there is nothing to catch the moment a bishop is quietly teleported or a pinned piece "moves" anyway. Nothing in next-token prediction itself rewards a legal move over a fluent-looking one; it rewards text that reads like chess. That's fine for describing a game and fatal for playing one.

What is ChatGPT actually good at in chess?

It's good at explaining, not calculating. Ask it to name the idea behind the Italian Game, walk through why a fork wins material, or talk you through a plan in a fixed position, and it draws on an enormous library of chess prose to teach clearly. The strengths are real and narrow: opening ideas rather than deep lines, tactical motifs (pins, forks, skewers, back-rank mates), and drilling positions where the correct answer is already known and can't be argued with. That split — the model explains, the card remembers — runs through every game in the library.

How do you get honest chess practice from an AI?

You add a referee — a real engine, or a pre-written answer key the AI grades against and can't overrule — so the coaching stays sound even when the model's confidence outruns its accuracy. That's the design behind Counterplay: every puzzle ships with its position and its solution written on the card, hand-built and then checked against the Stockfish engine — each solution is the engine's best move or a forced mate — so the AI never computes chess; it reads the answer and helps you find it. Eight puzzles, CP-1 through CP-8, easy to hard: a back-rank mate, Philidor's smothered mate, a pawn push that forks two knights. The openings trainer walks the Italian, the Ruy Lopez, and what White is really buying with the Queen's Gambit.

The card says it out loud: "I'm a coach and puzzle-master, not a chess engine. I can't reliably hold a full board in my head — I'll eventually make an illegal move or a blunder."

There's no cryptographic seal on it, the way there is on the sealed mysteries — no surprise to protect here, just an answer the coach isn't allowed to argue with. It shares the chess and strategy shelf with The Resumption, seven textbook endgames whose verdicts were checked against the Syzygy endgame tablebases, and both keep the same house rule: use the AI for what it's genuinely good at, and let something else hold the answer.

It's free and needs no signup: open the card, use the play link or paste the whole thing into a fresh ChatGPT, Claude, or Gemini chat, and say puzzles. Stop whenever — ask it to save, and the block picks up in any assistant (how saves work).