The state of AI in chess
In short Engine strength is no longer what holds an improver back. Stockfish 18 arrived in January 2026 and is free, and anyone with a browser can get deep cloud analysis. The new money and attention have gone into wrapping engines in coaching: voiced coaches on Chess.com, human-like models such as Maia that show what players at your rating actually play, and AI hints in Fritz. None of these has published evidence that it raises anyone's rating. General-purpose chatbots are still poor players. The best of them now rarely make illegal moves, but they cannot be trusted to explain a position without an engine checking them. Meanwhile, cheat detection keeps scaling up, and FIDE's new president has floated AI-backed identity checks for online play.
What it found
- The strongest engines got stronger again: Stockfish 18 claims up to 46 Elo over Stockfish 17, and Fritz 21 is advertised at 3,620, so the engine itself is not what limits a club player's analysis.
- General-purpose language models remain weak and unreliable chess players: in Google DeepMind's Game Arena study some models needed more than five forced retries per game for illegal or malformed moves, and GPT-4o lost to a 1979 Atari cartridge on beginner level.
- The major platforms are putting their AI effort into coaching layers, such as Chess.com's Play Coach and its cloned celebrity voices, rather than stronger analysis, and none of these products has published evidence that it improves results.
- Models that predict human moves are maturing: Maia 2 powers a free analysis and training platform, and the Chessformer paper reports 57.1% human move-matching with under a quarter of the previous best model's parameters.
- Online cheat enforcement is growing and increasingly automated: Chess.com closed 458.2K accounts in Q2 2026 with a 93.3% automated closure rate, and Lichess took 191,508 fair play actions in 2025.
- Chess governance is starting to engage with AI: Timur Turlov, elected FIDE president in September 2026, campaigned on an AI-powered system to verify who is behind online accounts.
How it was made
Every development in the report comes from a public source dated from May 2025 to September 2026: trade press, regulators, journals and company announcements. Each one is dated and names its source, so you can read the original, and every link was checked when this edition was put together. It is a record of what happened, not advice.
A new edition replaces this one as things move. The latest is always on this page.