Why Chess Bots Play Five Good Moves, Then Hang a Queen

A chess bot that plays five accurate, well-known moves and then hangs its queen for nothing is not simulating a human opponent, it is exposing how it was built. Real people who make that mistake usually make it for a reason a coach could name: a pin they missed, a check they never calculated, a plan they committed to too early. A bot that blunders that way because its search depth got dialed down for one move is a different thing entirely, and players notice the difference even when they can't name the mechanism.

The pattern is real, and chess.com's own users describe it

On the Chess.com community forum, in a thread titled "Maia Chess," user PsychoPanda13 wrote: "The chess.com bots will play 5 lines of straight opening theory and then just play one move which makes no sense whatsoever," adding that "their moves just 'feel' ridiculous and my games are nothing like those against real people." Another user, NikkiLikeChikki, put it more bluntly: "The bots know WAY too much theory for their rating."

Those are forum posts, not a controlled study, but the complaint is specific and consistent: a bot that opens with book-perfect theory and then produces a move with no plan behind it, rather than the steady stream of small, explainable errors a human at that rating actually makes.

Why depth-limited engines blunder like that

The mechanism isn't specific to any one product, and it has been described plainly. A Stanford CS224R course project report on modeling human-like play, an unreviewed student write-up rather than a peer-reviewed paper, describes the standard method of weakening an engine, cutting its search depth or randomizing among near-top candidate moves, this way: it produces weaker opponents whose play "alternate[s] unpredictably between brilliant strategic moves and inexplicable blunders, fundamentally unlike the systematic, consistent errors characteristic of human players at specific ELO ratings."

That's the mechanism in one sentence. A full-strength engine given a shallow search horizon still finds the objectively strongest move in most positions, because most positions have one move that's clearly better than the rest even at low depth. Occasionally, though, a position needs deeper calculation to avoid a trap. The shallow search misses it, and the bot drops a piece it would never drop with its normal depth restored. The result is bimodal: strong, strong, strong, then a random-looking catastrophe, not the graded, positionally motivated mistakes of a person actually thinking through the position.

That's a fair description of naive depth-limited bots, and of many of the simple free online bots that use them. It is not a fair description of Chess.com's current bot roster, which runs on the Komodo engine across more than 100 named personality bots (Martin, Nelson, Isabel, Aaron, Mittens at roughly 3200 Elo, plus celebrity tie-ins like MrBeast and Hikaru Nakamura), each with authored behavior rather than a single depth-limited engine reused at every level. A small number of those are free and the rest sit behind a paid Chess.com membership, a split Chess.com has adjusted more than once during 2026. Chess.com is also aware of the human-likeness problem and markets some of its own bots on exactly that basis. The forum complaints above are real, but the target should be naive depth-limited bots specifically, not a caricature of everything Chess.com ships today.

Hand-tuning a model to fake a psychological state doesn't obviously solve this either. One independent project, built around reshaping Maia-2's move probabilities to simulate states like stress or overconfidence, tested the result across 12,414 games. Move agreement with the plain Maia-2 baseline swung by roughly 20 to 25 percentage points depending on which artificial state was dialed in, evidence that small hand-built adjustments on top of a trained human-move model move the output a lot, and unpredictably. Human-like play isn't a knob you turn on top of any engine. It's closer to a property you have to train into the model from real games in the first place.

What actually predicts a human's next move

The credible fix comes from an actual line of published research, not a product feature. In 2020, researchers at the University of Toronto's Computational Social Science Lab, working with collaborators from Cornell University and Microsoft Research, published "Aligning Superhuman AI with Human Behavior: Chess as a Model System" (McIlroy-Young, Sen, Kleinberg, Anderson) at KDD 2020, posted as arXiv:2006.01855. The idea was different from building a weaker engine: train a model directly on how humans at a given rating actually move, using millions of real Lichess games as the training data, instead of starting from the objectively best move and degrading it.

That first version, now called Maia-1, was nine separate models, each trained on 12 million Lichess games and each specializing in a single 100-point rating band from 1100 to 1999. Tested against real human games, Maia's move-match accuracy ranged from about 46% in its worst case (a model trained on 1900-rated games predicting what an 1100 would play) up to 52% or more when predicting players inside its own training band. Attenuated Stockfish and Leela configurations, tested the same way as human-move predictors rather than as the strongest-move finders they're built to be, topped out around 35 to 40%. Maia's floor beat the depth-limited engines' ceiling.

Four years later, at NeurIPS 2024, the same research lineage published Maia-2, "A Unified Model for Human-AI Alignment in Chess" (Tang, Jiao, McIlroy-Young, Kleinberg, Sen, Anderson), arXiv:2409.20553. It replaced the nine separate rating-bin models with a single network using a skill-aware attention mechanism that integrates player strength directly with the encoded position, so one model covers the whole spectrum instead of needing a model swap every 100 rating points.

The newest version, Maia-3, is built on an architecture called Chessformer: an encoder-only transformer that treats board squares as tokens, adds a chess-specific positional encoding called Geometric Attention Bias, and predicts moves through an attention-based source-destination move head rather than a fixed output layer. It was published as a conference paper at ICLR 2026, arXiv:2605.19091. The reported headline number is that the 79-million-parameter Maia-3 checkpoint hits 57.1% move-matching accuracy, beating a prior 355-million-parameter model that used search at 55.9%, a smaller, purpose-built model outperforming a much larger general one with search bolted on.

Maia-3 ships three official preset checkpoints: maia3-5m at 5 million parameters, maia3-23m at 23 million, and maia3-79m at 79 million, plus a smaller 3-million-parameter research ablation. All of them target the same 600 to 2600 rating range on the Lichess scale, the range Maia's own site describes as covering any rating level on the Lichess.org rating scale.

Maia is free right now, and you can try it in five minutes

Maia is free and open source. Its makers describe Maia-3 explicitly as free and open source, and you can play it today with nothing but a free Lichess account: no download, no app, no payment. Lichess hosts the models as bot accounts. BOT maia1 and BOT maia9 run the original Maia-1 rating-band models, and newer community-run accounts under the maia3-79m name run the Maia-3 architecture at fixed target ratings. These are third-party accounts that can be renamed or taken offline without notice, so search "maia3-79m" in Lichess's bot directory for the current list rather than trusting a name printed here. Search for any of them on Lichess and challenge them directly. All bot play on Lichess is free.

That's a genuine point worth leading with rather than burying: if your only goal is to play against a research-grade human-like chess bot, Lichess already gives you that, at no cost, right now. Anyone selling "access to Maia" as the pitch is selling something you can get for free elsewhere.

It's also worth saying the same Chess.com forum thread that praised Maia's feel isn't uniformly glowing about it. One user flagged that a bot labeled 1100 played stronger than that number suggested; another described beating a supposedly rating-matched bot easily despite being rated lower themselves. Move-prediction accuracy and perceived difficulty aren't identical measurements, and Maia optimizes for the former, not for feeling exactly as strong as its label.

What a packaged app adds that Lichess doesn't

If Maia is free and open, the honest question is what a paid app adds at all. Not the model, that's the same research. What changes is the packaging around it.

Chessdrive, the app we build, bundles the actual maia3-5m checkpoint, the smallest of the three official Maia-3 presets, as a roughly 21MB ONNX file inside the app itself rather than downloading it or calling a server. It runs through ONNX Runtime entirely on the device, so a game against it works with no network connection and no Lichess account. Instead of picking a bot number or a rating-locked account, you set a single continuous ELO value, and that value conditions both Chessdrive's own play and its running model of what you, the opponent, are likely to do. The meaningful range is 600 to 2600, the same range Maia targets; the range surfaced in the app's own play setting is narrower, 1100 to 1900.

Stockfish does a completely separate job. Through the multistockfish package running at MultiPV=3, it evaluates the finished game afterward, not during it, so the opponent you played against and the engine that grades your moves are never the same system. Two engines, two jobs: one trained to predict what a human at your rating would actually play, one built to find the objectively correct move and show you how far you were from it.

Chessdrive is live on iOS and Android at $5.99 a month or $39.99 a year, with a 7-day trial on the annual plan and a 3-day trial on the monthly one. Being exact about our own limits, since this page has spent several paragraphs on what other people gate: the free install gives you three puzzles a day and a one-off credited game and review, not a recurring free allowance of games against Maia. If you want to play Maia repeatedly for nothing, Lichess is the honest answer and it is the one we pointed you at above. None of that makes Maia itself exclusive. It makes offline, integrated play-then-review a single flow instead of two separate tools.

How to try each option today

For pure access to Maia at no cost: Lichess, searching BOT maia1, BOT maia9, or one of the maia3-79m accounts, is the fastest path and needs only a free account.

For a large roster of authored, named personalities rather than a single skill dial: Chess.com's computer-play page runs its Komodo-based bots, with a small number free and the rest behind a paid membership. Chessiverse takes a similar personality-bot approach at larger scale, marketing itself directly around 1,000-plus bots spanning 100 to 3,300 Elo.

For an on-device, offline opponent paired with a separate post-game engine review in one app: that's the gap Chessdrive is built to fill, and it's the honest place a packaged product earns its price. Not by owning the model, but by owning the flow around it.

Last verified: 28 August 2026, against the sources cited above. Model checkpoints, Lichess bot accounts, and Chess.com's bot roster are all subject to change; if you're reading this months later, confirm current details before relying on them.