Guide

Human-like chess bots: what Maia does differently

By the Best Chess Bot team (ChessHelper) · Published November 12, 2026 · Updated November 12, 2026

Most chess bots are strong engines with the strength turned down. Human-like bots are built the other way round: trained on millions of real games to play the moves people actually play. This guide explains the research behind them, how they differ from low Stockfish levels, where you can play them, and how to review those games with ChessHelper (chesshelper.ai), the tool our team makes.

Why a weakened engine feels robotic

The usual way to make an engine easier is to limit it. Lichess's Stockfish levels, for example, combine three limits, according to the open-source fishnet code that runs them: a skill setting (from -9 at level 1 up to 20 at level 8), a maximum search depth (5 plies, or half-moves, for levels 1 to 5) and a time cap (50 milliseconds at level 1, one second at level 8). Stockfish's skill setting works by sometimes choosing a weaker candidate move instead of the best one.

The result is an opponent with an odd rhythm. It may defend perfectly for several moves, then drop a rook in a quiet position where no person would. It rarely makes the typical human mistakes: overlooking a backward knight move, grabbing a poisoned pawn, or attacking too early. Those are the errors you need to practice against, because they are the ones your real opponents make, and the ones you make yourself.

The Maia research project

Maia comes from a research team at the University of Toronto, Cornell University and Microsoft Research. The first paper, "Aligning Superhuman AI with Human Behavior: Chess as a Model System" by Reid McIlroy-Young, Siddhartha Sen, Jon Kleinberg and Ashton Anderson, was presented at the ACM SIGKDD conference (KDD 2020). Its question was simple: can an AI predict the move a human at a given rating will play, not the best move?

The team trained nine models, one for each 100-point rating band from 1100 to 1900, on roughly 12 million Lichess games per band. The paper reports these move-matching results on held-out games:

That last row is the key finding for anyone choosing a bot. The paper found that shallow Stockfish is not playing moves that are more typical of weak players; it is simply a worse engine. Maia, by contrast, captures how players at a given level actually think, including their mistakes.

A follow-up paper, "Maia-2: A Unified Model for Human-AI Alignment in Chess" by Zhenwei Tang, Difan Jiao, Reid McIlroy-Young, Jon Kleinberg, Siddhartha Sen and Ashton Anderson (NeurIPS 2024), replaced the separate per-band models with one model that takes the player's skill as an input. The authors report that it predicts human moves more accurately than the original Maia and behaves more consistently as skill changes.

A human-shaped mistake: the Blackburne Shilling trap

Here is the kind of position human-like bots are good at reproducing. After 1. e4 e5 2. Nf3 Nc6 3. Bc4 Nd4, Black's knight jump is objectively dubious; Stockfish rates White about a pawn and a half better after 4. O-O or 4. Nxd4. But it sets a trap that catches club players. White takes the pawn with 4. Nxe5, and Black replies 4... Qg5, attacking both the knight and g2.

All moves and evaluations in this example were checked with Stockfish at depth 18.

Where to play human-like bots in 2026

Maia bots on Lichess

The research team runs three bot accounts on Lichess: maia1, maia5 and maia9. Their profiles say each was trained on over 10 million Lichess games between players rated around 1100, 1500 and 1900 respectively. They are free. Their Lichess ratings move with every game and can sit some distance from their training band, so judge them by how the games feel, not just the number.

The Maia website

The team also runs maiachess.com, where you can play Maia at a chosen rating level and use extra tools such as game analysis, opening drills and a Hand and Brain mode. Details change as the project develops, so check the site for the current options.

Chessiverse

Chessiverse is a commercial platform with over 1,000 bots that, according to its own site, are built on Maia neural nets with its own adjustments, calibrated across a 0 to 3300 range. Free accounts get a curated selection and a paid Premium plan unlocks the full roster.

Chess.com personality bots

Chess.com's 100+ bots are all powered by the Komodo engine, according to its help center. They have distinct personalities, openings and chat, which makes them fun, but they are tuned engines rather than models trained to imitate human move choices. Its adaptive bots, which shift strength during a game, are a useful middle ground.

Which kind of bot should you use?

Use a human-like bot when you want practice that transfers to games against people: realistic openings, human-style errors to punish and positions that look like your own games. Use a strong conventional engine when you need the truth about a position, such as checking an opening line or practicing a winning endgame. Use a weak conventional engine mostly at the very start, when any opponent that lets you finish a game is useful.

Whichever you choose, the improvement comes after the game. Human-like bots produce human-like positions, so reviewing them with suggestions pitched at your level is especially productive. ChessHelper lets you set its engine strength from 800 to 3200 Elo for that review. Our learning from bot games guide walks through the routine, and chess bots by rating helps you pick the right Maia model.