Magnus Becomes Latest GM To Take On AI In A Game Of (Sort-Of) Chess
GM Magnus Carlsen, following in the footsteps of GMs Praggnanandhaa Rameshbabu and Hikaru Nakamura, played a blindfold chess game against ChatGPT in a video published Tuesday. In less than 48 hours, it had already racked up more than 100,000 views.
From move one, some of the things that make Magnus Magnus were evident. While both Pragg (video here) and Hikaru (link) led with 1.e4, Magnus chose a different path (0:52) and played 1.h4. Neural network chess engines such as AlphaZero—which, unlike ChatGPT, can actually play chess—were at the forefront of several ideas that are now a core part of chess. One such idea is the aggressive advance of the h-pawn... but not usually on move one!
If you remember some of our 2024 articles of ChatGPT giving really bad chess advice and taking on Gemini one-on-one, you may be surprised at how long GPT held its own in its game against the world's best human chess player. Stockfish, in fact, slightly preferred Black's position for several moves after Magnus's intentionally dubious opener.
Unfortunately for GPT—but fortunately for us—once it did err (3:23), all heck broke loose. Asked to explain itself, the computer seemed to lose track of a knight, forgetting it had played the steed from d7 to b6, and the lapse cost the silicon its queen.
GPT's first illegal move soon followed, but here again Magnus played along, allowing it to ignore an absolute pin and hang its king. The chaos only continued, with Magnus also making illegal moves—his, of course, on purpose. After moving his own king eight squares at the same time and then jumping it with his queen (that's how the knight moves, Magnus—we thought you knew this) to reach a checkmating position, Carlsen finally put his foot down when GPT attempted to continue the game by capturing the queen with yet another disallowed move.
Keep in mind Magnus tracked all of this without so much as a glance at the actual position! You do not want to miss this video.
The no-contest was the first time Magnus had played ChatGPT on video, but not the first time ever. Perhaps this explains him taking a different approach to the game than Pragg or Hikaru did... even if, in the end, GPT undershot its estimate of Carlsen's strength as a player by about 1000 Elo.
I sometimes get bored while travelling pic.twitter.com/MmkEeCg0Xn
— Magnus Carlsen (@MagnusCarlsen) July 11, 2025
These games are part of a larger trend of seeing how far large language models (LLMs) can go in chess, including the recent Kaggle Game Arena, which was actually won by the o3 version of ChatGPT, ahead of Grok 4 and six other AI models. Magnus himself even commentated, along with GM David Howell, for the final GPT-Grok match.
Commentating the @kaggle final together with an above average english chess player on the @TakeTakeTakeApp app.
— Magnus Carlsen (@MagnusCarlsen) August 7, 2025
Tonight @ 19:30 CEST pic.twitter.com/odbxkytMVv
It remains to be seen if the LLMs can eventually compete with humans in chess, let alone with the traditional, chess-dedicated engines which long ago left humans in the dust. Cutting out the illegal moves would be a start, but after Hikaru's turn taking on GPT, he was pessimistic about the LLMs' chances (4:14).
For now, watching LLMs try to play chess is an entertaining sideshow from the serious side of artificial intelligence as it continues to change the world (for better and for worse). Be on the lookout for more GM vs. AI battles in the future!
What was your favorite move in Magnus's blindfold game against ChatGPT? Let us know in the comments!
Nathaniel Green is a staff writer for Chess.com who writes articles, player biographies, Titled Tuesday reports, video scripts, and more. He has been playing chess for about 30 years and resides near Washington, DC, USA.
