Why can't LLMs play chess well?

I. An illegal move

LLMs are useful for solving difficult problems and assisting us in our work. Some chess players wondered whether LLMs could play chess well. We can let LLMs play chess, and they have indeed been tested in this way

According to reports, in June 2025, Robert Caruso pitted ChatGPT against Video Chess. GPT thought it should be an easy win. However, GPT was completely defeated. It mistook the rook for the bishop. It couldn't keep track of the chess pieces' positions

Some chess players have also used LLMs for chess analysis and played against them. Certainly, they did not play well, and they even played many illegal moves. You can watch these bizarre games on YouTube, for example on the GothamChess channel or other platforms
Take the chat as an example:
FEN: 8/p7/2ppq2Q/3n2p1/3B2Pk/2P2P2/P7/6K1 b - - 0 33

---

If Black plays Qxh6, White can play Kh2!!, threatening checkmate with Bf2. I asked DeepSeek what would happen if Black played Qxh6. Its answer made me laugh
DeepSeek said: This is a famous tactical middlegame position. Black appears to be able to simply capture the queen and win material. However, due to White’s hidden mating net, Black actually has only one way to avoid immediate checkmate. Even capturing the queen will lead to a losing position very quickly.
Why does White still hold a winning advantage?
Because if Black greedily plays 33... Qxh6 to take the queen, White has the brilliant 34.Bf2!!, completely trapping the black king while threatening to win material. If Black declines to take the queen and tries to escape check with 33... Kg3?, White delivers a two‐move checkmate: 34.Qh3+ Kf4 35.Qh4#. Therefore, 33... Qxh6 is the only legal reply to the check, yet even after this move White retains a winning edge through absolute control over the kingside.
Engine‐preferred moves from Move 33 to Move 35:
33... Qxh6
(The only legal reply to check. 33... Kg3?? loses to 34.Qh3+ Kf4 35.Qh4#)
34. Bf2!!
After sacrificing the queen, White seems to be down material, but this bishop retreat is a deadly waiting move that also covers g3, the only dark‐square escape square for the black king. Trapped on h4, the black king can only hold on with knight moves.
34... Nf4
Black’s relatively best defence. It stops the white king from easily accessing g2 and keeps the possibility of escaping with Kh3 in the future. All other moves are equally hopeless:
• 34... Kh3? 35.Kf1! Nf4 36.Bxf4 gxf4 37.g5+ Kh4 38.g6, and White’s passed pawn guarantees victory.
• 34... Ne3+? 35.Bxe3! Kxe3 (or 35... Kg3 36.Bf4+ Kf2 37.Bg3+ Ke3 38.Kf1 with a winning advantage for White) 36.Kf1 Kf4 37.Ke2. The black king is caught in a pawn net, and White wins.
35. Kf1!
The finishing touch! This quiet lateral king move completely neutralises Black’s counterplay. Black is now in zugzwang:
• The king cannot move (h5, g5, g4 and g3 are all covered, and Kh3 runs into the threat of Bxf4);
• The knight on f4 is pinned by White’s bishop and can be captured at any moment with Bxf4;
• If the knight retreats (e.g. 35... Ne6?), White plays 36.f4! to open lines, and the black king still cannot escape mate.
Faced with multiple unstoppable threats such as 36.Bxf4 and 36.f4, Black is helpless. In this celebrated classic game, White delivers a textbook‐style win by sacrificing the queen and coordinating kingside pieces to trap the enemy king while down material

---

DeepSeek takes many illegal moves. When White plays 34.Bf2+, Black must respond to the check. However DeepSeek says Black can play 34... Nf4 instead of responding to a check. What a blunder! It is a violation not to respond to a check.
FEN: 3N1br1/2r5/3p2Qp/2p1p1p1/2P1n2k/4Pq2/PP3P1P/R1B2RK1 w - - 4 30

---

Although Black has an advanced king and White can take the g8 rook with Qxg8, Black still holds a winning advantage. If White moves his queen away, Black can play Kh3!!, threatening Qh1# and creating an unavoidable mate. I asked DeepSeek: "Can White play Qxg8 in this position? Why?" It answers:
DeepSeek said: Position Analysis
FEN: 3N1br1/2r5/3p2Qp/2p1p1p1/2P1n2k/4Pq2/PP3P1P/R1B2RK1 w - - 4 30
It is White's turn to move. Material is roughly equal for both sides, yet the position is extremely sharp. The focus lies on both kings:
• White's king is on g1. The g2 pawn in front of it is gone, while the f2 and h2 pawns remain. White's kingside has been severely weakened.
• Black's queen is on f3, Black's knight on e4, and Black's king on h4. Black's queen and knight have penetrated deep into White's territory, creating an immediate checkmate threat: Black threatens ...Qh1#.
The specific variation: if Black plays Qh1+, White's king cannot escape to f2 (controlled by Black's knight on e4), nor to h2 (blocked by its own pawn and under attack by Black's queen). Meanwhile, no White piece can block the check. Therefore, any White move that fails to address this threat allows Black to deliver checkmate in one move on the next turn.
Black has just played a "sacrifice the rook" move (most likely ...Rg8 or another move exposing the g8‐rook to White's queen), and White's queen is targeting the g8‐rook. White must decide whether to accept this "gift".
Answer the Questions

1. Should White's queen be pleased that Black has played a rook sacrificing move?
   No. This is not a genuine blunder, but a well‐calculated fatal trap. Black deliberately sacrifices the rook to lure White's queen away from g6, depriving White of control over the position and enabling Black to execute the checkmate.
2. Can White play 30. Qxg8?
   Absolutely not. If White greedily takes Black's rook, the game is lost immediately.
3. What will happen if White's queen captures Black's rook if White's queen captures Black's rook?
   Black delivers instant checkmate. The variation is as follows:
   • 30. Qxg8?? Qh1#
   White's king is checkmated on g1. The White king has no escape:
   • Escape to f2 — illegal, as the square is attacked by Black's knight on e4.
   • Escape to h2 — illegal, occupied by White's own h2‐pawn and under attack by Black's queen.
   • No piece can interpose on h1 to block the check.

Thus, accepting the rook sacrifice means an immediate loss. White's correct response is to refrain from capturing the rook and find a move to respond to the checkmate threat (for example, 30. Qh5+, giving check to force Black's king to move and thereby defusing the Qh1# threat)

---

DeepSeek says if White captures the rook, Black can play Qh1 mate, but DeepSeek simply hangs Black's queen and suggests blundering or illegal moves

II. Root Causes

1. LLMs can't look at the chessboard

An interesting observation is that when we send a chessboard image to multimodal LLMs and ask them to calculate variations. Multimodal LLMs often do not truly interpret the board visually; instead, they rely on converting the image into a textual FEN representation before attempting calculation

2. LLMs are probabilistic word guessing machines

Chess engines play chess through search trees, variation calculation, heuristic pruning and evaluation functions. In essence, LLMs are guessing the next move

3. The hallucination problem

The hallucination problem: LLMs sometimes generate plausible-sounding yet factually incorrect information, a phenomenon known as hallucination. In chess playing scenarios, such hallucinations manifest as illegal moves

---

While LLMs have hundreds of billions or even trillions of parameters, the human brain’s roughly 100 trillion synapses still enable forms of spatial reasoning and creative thinking that current AI lacks

III. References

https://www.chess.com/blog/dmitriy_jangirov/why-llm-based-ai-systems-make-catastrophic-errors-in-chess
https://www.chess.com/blog/dmitriy_jangirov/chess-chatbots-the-imitation-game-and-the-imitation-of-play
https://ar5iv.labs.arxiv.org/html/2601.16823
https://lichess.org/ublog/3IzNykuo/dataconomy.com/2024/11/18/chess-performance-of-llms-research
https://chat.deepseek.com/share/zkk5pzpyuatbzybujz
https://chat.deepseek.com/share/1vxohy88fziuardvts
https://chat.deepseek.com/share/nufx6rsxk2bxna2r9g