It was meant to be a fun exhibition game between the world’s greatest human chess player and one of the world’s most talked-about AIs. But when Magnus Carlsen played ChatGPT blindfolded, the match quickly spiraled from logic to lunacy, culminating in both sides making illegal moves, and Carlsen checkmating the AI with a king that moved like a knight.
Opening Fireworks and ChatGPT’s “Human” Thinking
From the start, Carlsen’s play was aggressive but unconventional. His first move? h4, the kind of flank pawn push that makes traditional chess engines raise an eyebrow. ChatGPT responded with d5, setting the tone for a strange but energetic opening.
Carlsen followed with d4, e3, and f3, hinting at an ambitious pawn center. ChatGPT pushed back with logical if slightly passive moves like e6 and c6, developing steadily into a queen’s pawn structure.
As the game progressed, it became clear that ChatGPT’s style was not the cold, tactical precision of Stockfish or Leela Chess Zero. Instead, its moves — good or bad — had a distinctly human flavor. Sometimes it made strong, thematic plays. Other times, it hallucinated threats, misread pins, or confused knight positions.
Carlsen later said this was precisely what fascinated him about LLMs playing chess:
“Their mistakes and the interpretations that led to their mistakes were both funny and educational,” Carlsen told Take Take Take. “They’re like ambitious club players, not perfect, but full of ideas.”
The Middle Game: Pins, Sacrifices, and Confusion on the D-File
Midway through the game, Carlsen castled queenside, which ChatGPT immediately criticized as “risky” due to a loose pawn structure. But when asked if it saw the danger along the D-file, the AI became confused, claiming that a pinned knight on d7 was the problem, even though there was no such pin.
Carlsen exploited this foggy reasoning with a sharp Ne6 sacrifice. ChatGPT took the knight with fxe6, but Carlsen replied with Qxe6, cranking up the pressure. What followed was a flurry of exchanges, with rooks and bishops disappearing from the board.
By this stage, ChatGPT had avoided outright illegal moves, but not without some suspicious “thoughts.” At one point, it announced it wanted to play Rf4, a move that wasn’t legal in the position. The human operators ignored it and entered a valid move instead.
Descent into Madness: Illegal Moves Everywhere
And then… the chessboard reality broke down.
Carlsen, sensing the absurdity of the moment, advanced his king deep into enemy territory with Ke6. ChatGPT praised the move for “attacking the rook on f8” and then, in an apparent glitch of chess geometry, claimed the king was moving “like a knight” to put pressure on the position.
From there, the moves turned surreal. Rooks began appearing on strange squares, pawns seemed to reanimate from the grave, and piece captures defied the laws of chess physics. Carlsen laughed as the game crossed into full exhibition comedy.
The final straw came when Carlsen delivered checkmate using a queen move that simply wasn’t legal. The audience (and even ChatGPT’s handlers) acknowledged that the game had long since abandoned the rules. The scoreboard read: Magnus Carlsen 1 – ChatGPT 0.
Bigger Picture: AI Chess, Human Style
Despite the chaos, the exhibition revealed something important: LLMs like ChatGPT think about chess in ways that resemble human club players, not in the hyper-accurate, brute-force style of traditional chess engines.
Carlsen has praised this quality before, noting that such AIs play at a level roughly below 2000 Elo, strong enough to challenge casual players, but full of instructive mistakes. This is a far cry from machines like Stockfish, which sits at an astronomical 3731 Elo and is utterly untouchable for humans.
“If we consider LLMs as relatively new players, we need to encourage good ideas, even if they’re wrong,” Carlsen said. “Their games are fascinating to watch.”
ChatGPT’s Competitive Record
In fact, ChatGPT is no stranger to success in its own league. Just before this match, its o3 model won the first-ever AI Chess Championship, defeating Elon Musk’s Grok 4 by a clean 4–0 score in the finals. Across the tournament, it averaged 91% move accuracy, impressive, though still far from grandmaster strength.
The tournament was unique in that it featured non-chess-specialist AIs, unlike traditional chess engines. That meant their games were unpredictable, sometimes creative, and occasionally baffling, much like the Carlsen match.
Blindfold Brilliance… and a Wink at the Audience
For Carlsen, this wasn’t about proving superiority. He’s already done that countless times against both humans and machines. This was about fun, curiosity, and showing that even the best AI brains can get flustered over a simple pin or miscount a square.
And maybe, just maybe, it was about reminding everyone that chess is a game before it’s a science, a place where you can laugh at a “knight-moving king” and still deliver checkmate.
As Carlsen put it with a grin after the game:
“It almost lasted 30 moves without making an illegal move. That’s something we can all be proud of.”
I’m a passionate board game enthusiast and a skilled player in chess, xiangqi and Go. Words for Attacking Chess since 2023. Ping me at Lichess for a game or chat.
