Maia is a chess neural network trained to predict the move a human would play, not the best move. Where an engine like Stockfish searches for the objectively strongest continuation, Maia learned from millions of online games between real players and answers a different question: "what would a 1500-rated player do here?" The result is an opponent that plays — and errs — like a person at a chosen rating, which makes it the most useful sparring partner I know of for club players. It was created by university researchers, its second generation (Maia-2) handles every rating level in a single model, and you can play against it on ChessMind AI at six strengths from about 1100 to 2000+ Elo. This guide explains how it works, why it beats a weakened engine as a training partner, and how I recommend using it.
Engines Play Perfect Chess; Maia Plays Human Chess
Every chess engine you have ever played against was built to win. Stockfish evaluates tens of millions of positions per second and picks the move with the best score; its "easy" levels are the same machine with a handicap bolted on — less search depth, or a random blunder injected every few moves. That is why playing a weakened engine feels so strange. It defends a lost position with inhuman precision for twenty moves, then drops a rook for no reason. It never gets nervous, never misjudges a quiet position, never falls for the trap that every club player falls for. You are not practicing against a 1500 player; you are practicing against a grandmaster with a random-number generator.
Maia was built the other way round. Instead of searching for good moves, it was trained to imitate moves — to look at a position and predict what a player of a given rating actually chose in it. Its teacher was not chess theory but the public game database of Lichess, millions of games played by ordinary people at every level. So Maia's mistakes are the mistakes humans make: it misses the quiet retreat, overvalues a pawn, walks into a pin, rushes an attack that isn't ready. When it is "weak", it is weak in the same places you are.
Where Maia Comes From
Maia is a research project, not a commercial product. The first version was published in 2020 by researchers from the University of Toronto, Cornell University and Microsoft Research, in a paper about aligning superhuman AI with human behavior — chess was their test case because it has both superhuman engines and enormous archives of human games. That first generation was a family of separate networks, one per rating band from 1100 to 1900, and each one matched the human move far more often than any engine did.
The second generation, Maia-2, was presented in 2024. Its key change is that a single model covers every skill level: you tell it the rating of the player and of the opponent, and it predicts the move for that matchup. That makes its behavior consistent across levels — the 1300 and the 1700 versions are the same "mind" at different strengths, rather than two unrelated networks — and it is the version ChessMind AI runs.
How Maia Picks a Move
It helps to understand the mechanism, because it explains both Maia's strengths and its limits.
Given a position and a target rating, Maia produces a probability for every legal move: how likely a player of that rating is to choose it. In a simple recapture almost all the probability sits on one move, and Maia plays it like anyone would. In a rich middlegame the probability is spread out — the strong move, the natural-looking move, the tempting mistake — and the spread depends on the rating. At 1100 the tempting mistake carries a lot of weight; at 1900 it carries much less.
Two consequences follow. First, Maia does not calculate deeply the way an engine does, so its errors are judgement errors and short-sightedness errors — the human kind — rather than random collapses. Second, it is not trying to beat you. A human-like opponent at your level will make the move your opponents make, which is exactly what you need to rehearse against. If you want the best move in a position, you still want an engine; my guide on how to analyze your chess games shows where Stockfish belongs in your training.
Why a Human-Like Opponent Is Better Training
Players who train mostly against engines tend to learn to play against engines. They expect perfect defense, so they never learn to set problems. They see random blunders, so they learn to wait for gifts. Neither habit survives contact with a real opponent.
A human-like sparring partner fixes three things at once.
You practice converting real advantages. Against Maia at your level, a better position stays better only if you keep making good moves — it will not hand you the game, and it will not defend like a machine either. That is the conversion skill that separates club ratings, and you cannot rehearse it against an opponent who is either flawless or random.
You meet the openings and plans of your real opponents. Because Maia learned from human games, it plays the main lines and the popular sidelines that people at your rating play, and it reacts to your opening the way they do. Rehearsing a repertoire against it is much closer to a tournament than drilling lines against an engine's first choice.
You get an honest difficulty dial. Set Maia about 100–200 points above your rating and the games are hard but winnable — the zone where you improve fastest. It is the same principle I describe in how to get better at chess: train at the edge of your ability, with feedback, on the thing that decides your games.
How to Train With Maia
Here is a routine that works.
- Pick a level just above yours. If you are rated around 1400 online, start at the ~1500 level. Move up when you win more than you lose over twenty games; never move up after one good evening.
- Play slower games. Maia is most instructive at rapid time controls, where you have time to think and the result reflects your decisions rather than your mouse speed.
- Review every loss. Find the move where the evaluation turned and ask what kind of mistake it was — a missed tactic, a bad plan, a lost endgame. Human-like opponents produce human-like turning points, so the review teaches you about your real games.
- Rehearse your openings against it. Play your repertoire against a human opponent until the typical middlegame positions feel familiar, not just the first ten moves.
- Practice positions, not just games. The most efficient work is to start from a position you need to learn to handle — an endgame, a typical structure — and play it out against human-like resistance.
Where You Can Play Against Maia
The research team publishes Maia's code and runs Maia bots on Lichess, which is the easiest way to try the original models.
On ChessMind AI, Maia-2 is built into the training rather than offered as a separate curiosity:
- Play out any position from the training — a middlegame exercise or a daily lesson — with its Play this position button, against one of six levels — Beginner (~1100), Casual (~1300), Intermediate (~1500), Advanced (~1700), Expert (~1900) and Master (2000+) — and download the PGN of the game afterwards.
- Practice your openings in the interactive opening courses: play mode starts a game against Maia from the course's opening position, so you rehearse your repertoire — and the middlegames it leads to — against the moves human players actually choose.
What Maia Is Not
To keep expectations honest: Maia is not a coach. It does not explain its moves, and the move it plays is the likely move, not the right one, so never use it to decide what you should have played — that is the engine's job. It is also not a perfect imitation of any single person; it imitates the average player at a rating, so it is more consistent than a real opponent who has good and bad days. And at the very top levels, human play becomes harder to predict, so Maia's human-likeness is strongest in the club range where most of us live.
Used for what it is — a realistic opponent at a level you choose — it is the best sparring partner chess has ever had for the ordinary player.
Frequently Asked Questions
What is Maia chess?
Maia is a neural network trained on millions of online games to predict the move a human player of a given rating would make, rather than the best move. It was created by researchers from the University of Toronto, Cornell University and Microsoft Research, and plays like a real person at a chosen level, mistakes included.
How is Maia different from Stockfish?
Stockfish searches for the objectively best move and plays at superhuman strength; its easy levels add artificial handicaps such as random blunders. Maia imitates human decisions at a specific rating, so it makes the natural, human kind of mistakes. Use Stockfish to analyze positions and Maia to practice against a realistic opponent.
What is Maia-2?
Maia-2 is the second generation of Maia, presented in 2024. Instead of one separate network per rating band, a single model covers every skill level and takes the ratings of both players into account, which makes its play more consistent from one level to the next. ChessMind AI's human-like opponents run on Maia-2.
Where can I play against Maia?
You can play the original Maia bots on Lichess, and Maia-2 on ChessMind AI, which offers six levels from about 1100 to 2000+ Elo. On ChessMind AI you play it in the opening courses' play mode, where the game continues from the course position, or from the positions in the training, with their Play this position button.
What rating should I set Maia to?
Set it about 100 to 200 points above your current rating, so games are hard but winnable. Move up a level only when you win more than you lose over a meaningful number of games, such as twenty, rather than after a single good session.
Is playing against Maia better than playing against people?
It complements human games rather than replacing them. Maia is always available, plays at a consistent level and lets you start from a course or training position, which makes it ideal for rehearsing openings and endgames. Serious games against people remain the real test, and the best routine combines both.