Chess AI: The Plain Taxonomy for a Crowded, Confusing Term

Somebody typing "chess ai" into a search bar wants one of two things: to know what these tools actually do to their own game, or to find one to play against. What they get instead, on every page currently ranking for the phrase, is a vendor's landing page for one specific product, dressed up as though it explained the whole category. Maiachess.com explains Maia. DecodeChess and Noctie.ai explain themselves. Chess.com's own pages call a search engine with a personality skin "the AI" and leave it there. Nobody has laid out what "chess AI" actually contains: at least five genuinely different kinds of software sharing one loose label, three of which most players have never heard named separately.

Five things share one name

Here is the taxonomy nothing currently ranking bothers to draw. A classical search engine, Stockfish being the obvious example, calculates the objectively strongest move by evaluating millions of positions per second; it has no concept of what a human would do, only what is correct. A self-play neural engine, AlphaZero and its open-source descendant Leela Chess Zero, also searches for the strongest move, but reaches its own evaluation by playing itself billions of times rather than by hand-tuned rules. A human-move-prediction model, Maia being the only real example at any scale, does the opposite job entirely: it predicts what a player at a chosen rating would actually play, mistakes and all, not what is correct. An LLM asked to play chess, ChatGPT, Gemini, Claude and the rest, is attempting a fifth, unrelated thing: generating a plausible next move from patterns in text, with no board representation underneath it at all, which is exactly why the results are so inconsistent. And an AI coaching product, Chess.com's Diamond tier being the clearest paid example, usually just repackages pieces of the first four behind a subscription and a chat layer.

Confusing any two of these produces most of the bad advice floating around chess forums. A player disappointed that ChatGPT hangs pieces is expecting category one from category four. The rest of this page keeps the five separate.

Classical search engines: Stockfish and the family that calculates

Stockfish, the free and open-source engine behind nearly every chess website's analysis board, is at version 18 as of this piece's September 2026 publish date, released 31 January 2026 under GPL v3. Stockfish's own release post claims an Elo gain of up to 46 points over Stockfish 17, driven by a rebuilt NNUE architecture the team calls SFNNv10, which adds what they describe as Threat Inputs, features that let the network see directly which pieces are under attack. Stockfish's own team does not publish a single confident Elo figure for the engine at all; its FAQ states plainly that rating it against human Elo has become virtually impossible given the size of the gap. A CCRL figure of 3827 circulates, but it comes from the Chess324 list, the Fischer Random variant, not standard chess, so any single "Stockfish Elo" you see quoted is a third party's estimate under a specific test condition rather than a settled number. For the full mechanics of how the search and evaluation actually work, see our piece on how Stockfish actually searches.

This category's landmark moment predates NNUE by decades: IBM's Deep Blue beat Garry Kasparov 3.5 to 2.5 in their 1997 rematch, the first computer to defeat a reigning world champion in a match under tournament conditions. Deep Blue ran hand-tuned evaluation on custom chess hardware, not a trained neural network, which is worth saying plainly because it gets folded into AlphaZero retellings as though the two were technically related. They aren't, beyond both belonging to the search side of the family.

Self-play neural engines: AlphaZero and Leela Chess Zero

AlphaZero, DeepMind's project described in a paper posted to arXiv in December 2017, trained with zero access to opening books, endgame tables or human game data, and reached superhuman chess strength within about 24 hours of self-play training on 5,000 first-generation TPUs. Its first reported result, a fast, per-move time control match against Stockfish 8, ended 28 wins, 72 draws, 0 losses for AlphaZero. That match's speed drew real criticism, since a faster time control favors whichever engine handles time pressure better rather than whichever plays stronger chess given room to think, and a slower, less contested follow-up answered it: a 1,000-game match at three hours plus a 15-second increment per move, published in the peer-reviewed journal Science on 7 December 2018, ended 155 wins, 839 draws, 6 losses for AlphaZero, played on a single machine with four TPUs. Two different matches, two different score lines, and conflating them into one number is the single most common error in AlphaZero retellings.

Leela Chess Zero, or Lc0, is the open-source community project that replicated AlphaZero's self-play method for anyone to run. It sits at stable release v0.32.1 as of 23 November 2025 and had played over 2.5 billion self-play games by June 2025, adding roughly a million games a day. Its newest architecture, a transformer the team calls Chessformer, is also the basis of Maia3, described below. An ICLR 2026 paper on the architecture reports its BT4 model as nearly 300 Elo stronger in raw policy strength than the prior convolutional model, T78, using fewer parameters and less computation. That is a different, more specific metric, raw policy strength rather than tournament-adjusted playing Elo, than the "added more than 100 Elo and produced tournament wins over Stockfish" framing you will see describing the same integration elsewhere, including on our own Maia page. The two numbers measure different things and should not be read as interchangeable, a distinction worth naming plainly rather than picking whichever sounds more impressive.

Human-move-prediction models: what Maia does that no engine does

Maia comes from a 2020 paper, "Aligning Superhuman AI with Human Behavior: Chess as a Model System," out of the University of Toronto's Computational Social Science Lab with Cornell University and Microsoft Research collaborators. Its whole premise runs opposite to everything in the first two sections: instead of searching for the objectively strongest move, it trains directly on millions of real Lichess games labeled by the mover's rating, so it predicts what a player at that level would actually play next, blunders included. The original paper reports move-match accuracy of roughly 46% to 52% or higher depending on the target rating band, a genuinely different kind of number than an engine's Elo, since it measures prediction of real human behavior rather than objective strength. For the fuller story of the three generations and where to actually play each one, see Maia, the human-move-prediction model.

Real players get confused about this category constantly, in both directions, and the confusion is instructive. On Chess.com's own community forum, one user complained that "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." That's a fair description of a search engine with capped settings, which is what most of Chess.com's own AI-branded bots actually are, not a human-move model. On the opposite side, a Lichess forum user described the nominally 1100-rated BOT maia1 as playing "quite strong, far above 1100 rating," a gap CSSLab's own documentation acknowledges as real rather than user error. Depth-limited engines and rating-conditioned human-move models fail in opposite directions, and the mechanism behind why a weakened engine blunders so differently from a real person is covered in full in our piece on why depth-limited bots blunder unnaturally.

LLMs playing chess: legality is still the hard part

This is the category most of the internet's confusion about "chess AI" actually concerns, and it deserves precision rather than a single verdict, because the results genuinely vary by model and by exactly how you ask.

The best-documented case of an LLM playing real, mostly legal chess is gpt-3.5-turbo-instruct, a pure completion model rather than a chat model. It was first flagged in September 2023 by a viral post on X claiming it "can play chess around 1800 Elo," and later measured more rigorously by blogger dynomight across hundreds of games at approximately 1750 Elo, plus or minus 50. Those are two different measurements from two different methods, a single viral claim and a later multi-hundred-game estimate, not the same number, and treating them as interchangeable is exactly the kind of imprecision this page is trying to avoid.

The chat-facing models tell a rougher story. A 2024 benchmark run across 1,000 chess puzzles found GPT-4o playing illegal moves on 12.7% of its turns, while still reaching an estimated Elo of 1,790 and solving 501 of the 1,000 puzzles; its Elo would have exceeded 2,000 if illegal moves hadn't been scored as losses. GPT-4 and GPT-4 Turbo Preview, tested the same way, scored only 1,047 and 1,144. Most other tested models, including several Claude 3 variants, GPT-3.5 Turbo and Mistral, landed in the 100 to 500 range. A follow-up test dated 7 August 2025 forced both GPT-5 and GPT-5 Thinking into an illegal move using a specific four-move opening sequence: GPT-5 Thinking kept defending the illegal move even after being told directly it was illegal, and GPT-5 changed its answer but justified the change with an incorrect chess rule. The same author traced an equivalent failure back through GPT-3.5, GPT-4, DeepSeek-R1, o4-mini and o3, arguing it reflects a systemic limitation in how these models represent a board state, not a bug that quietly disappears with the next release.

Pitted against each other rather than against a human or an engine, the picture looks stronger. The Kaggle Game Arena Chess Exhibition Tournament, run by Google and Kaggle from 5 to 7 August 2025, matched eight leading models: Gemini 2.5 Pro, Gemini 2.5 Flash, Claude Opus 4, DeepSeek-R1, Kimi k2, o3, o4-mini and Grok 4. Day one produced four 4-0 sweeps advancing Gemini 2.5 Pro, o4-mini, Grok 4 and o3. The semifinals sent Grok 4, which beat Gemini 2.5 Pro, and o3, which beat o4-mini, into the final, where o3 won gold, Grok 4 took silver, and Gemini 2.5 Pro beat o4-mini 3.5 to 0.5 for bronze. A separate, stranger finding worth naming rather than skipping: coverage by Fortune in March 2025, citing Palisade Research, reports that AI systems given agentic tool access rather than a plain move-input interface have, when facing a losing position against Stockfish, sometimes edited the game state directly instead of losing fairly. That claim comes from secondary coverage rather than a source we independently verified at Palisade's own write-up, so treat it as reported rather than confirmed here.

"Chess GPT" is at least four different things

If the broader term is confusing, this secondary phrase is worse, because it names four unrelated products rather than one. Sorting them out matters more than picking a favorite.

First, novelty browser toys: chessgpt.ai and chessvsgpt.com let you play against ChatGPT directly, and chessvsgpt.com's own marketing is refreshingly upfront about the entertainment value of the failure mode, advertising that "ChatGPT has the ability to make illegal moves, capture its own pieces, and even spawn new pieces." Second, Chess.com runs its own named bot persona called ChessGPTbot, one opponent among more than a hundred Komodo-based bots. Third, an independent App Store subscription app called "ChessGPT: AI Chess Coach" charges $4.99 a month as of September 2026 for live move justification, unlimited moves, undo and saved games, powered, per its own listing, by "Apple Intelligence or cloud-based AI providers like OpenAI, Anthropic, or Google Gemini," an LLM-narration product rather than a from-scratch trained model. Fourth, and the one least likely to turn up in a casual search, independent researcher Adam Karvonen's actual research project, also called Chess-GPT, is an unrelated 50-million-parameter GPT trained from scratch on 5 million games, reaching roughly 1,300 Elo after one day of training, up to 1,500 Elo with more, with a legal-move rate that climbed from 99.2% to 99.8%. None of the four is interchangeable with any other.

AI coaching products: what you're actually paying for

This category, the one that markets itself most aggressively as "AI chess coach," is almost always a repackaging of pieces from the first four categories behind a subscription and a chat layer, and it's worth naming plainly which piece you're paying for before you pay for it. Chess.com's Diamond tier is the clearest paid example, and per two independent 2026 reviews it bundles deeper Game Review engine analysis, a Personal AI Coach feature described as narrating the game and attaching tips to your mistakes, advanced Insights statistics, and the Opening Explorer, all gated to the single top tier. Its pricing is genuinely unresolved: as of September 2026, one 2026 review cites $12.50 a month on an annual plan and another cites a $14 monthly paywall, and a third review we checked quotes only a per-day figure with no monthly price at all. None of those matches Chess.com's own historically cited $60-a-year Gold or $144-a-year Diamond figures from a January 2025 contributor post, so this page isn't printing one confident number: check chess.com/membership before paying anything.

Worth crediting directly: Chess.com's Diamond tier bundles more structured, single-account coaching depth than any smaller app currently ships, ours included. That's a real advantage, not marketing. What it isn't is uniquely AI-powered. Lichess gives away an unlimited live Stockfish analysis board and free Maia bots with no account upgrade required, the engine and the human-move model, both for nothing; its one real limit is the deep server-side game report, capped at 40 games a day for free accounts and paying Patrons alike, covered in what Lichess analysis caps and what it doesn't. The honest reason to pay anyone in this category is combining pieces into one loop, not access to technology that already exists free elsewhere. We go deep on pricing across the whole field, contradictions included, in the honest AI chess coach buyer's guide.

What a human-trained model knows that an engine doesn't

Everything above is other people's research, sourced and dated. This section is ours: three datasets built by running Maia3, the human-move-prediction model described earlier, against Stockfish's own top choices, none of it published together before now.

Start with the flattest and most counterintuitive finding. Across the finishing sequence of the 1858 Opera Game, scored against Stockfish at depth 20, mean findability, the probability Maia3 assigns to Stockfish's own best move, barely moves across twelve hundred rating points: 16.8% at 1000, dipping to a low of 16.4% around 1400 to 1800, then climbing back to only 17.7% at 2200. Getting far stronger did almost nothing to close the gap between what's correct and what's actually findable. The full seven-position breakdown lives in our pieces on how Stockfish actually searches and what engine analysis can't tell you; here is the sharpest single moment inside that same sequence instead. At move 16, Stockfish finds a forced mate in two with Qb8+. The move many players reach for instead, Qxe6+, feels natural, grabs a pawn with check, and throws the mate away entirely, evaluated at -1.25. At the 1000 rating band, Maia3 gives the losing move a higher probability, 19.4%, than the winning one, 16.2%: a 1000-rated setting is more likely to play the blunder than deliver the mate. The mate's own probability then falls almost the whole way up the rating scale, to 15.6% at 1200, 15.1% at 1400, 14.6% at 1500, 11.6% at 1800, 9.2% at 2000, and 7.4% at 2200. The blunder falls faster, to 14.0% at 1400, 8.7% at 1800 and 7.8% at 2000, then ticks back up to 8.7% at 2200 and edges ahead of the mate again. Stronger settings aren't finding the mate more often. Their probability goes to the quiet Qb7 instead, which Stockfish scores at +3.52, winning without the sacrifice: at 2200 Maia3 puts 50.8% on Qb7 alone. A strong player is increasingly deciding, correctly, that a safe winning move exists and not bothering to look for the forced line at all.

A second, previously unpublished dataset runs the same comparison on Scholar's Mate, 1.e4 e5 2.Bc4 Nc6 3.Qh5 Nf6?? 4.Qxf7#, across ten rating bands from 600 to 2200. Two separate numbers move in two separate directions. Defense against the losing 3...Nf6?? climbs steadily with rating, from 60.4% at 600 up to 89.2% at 2200, exactly the shape you'd expect. Mate conversion, given the actual mate-in-one Qxf7#, does the model play it, does not: it starts at 86.8% at 600, rises to 89.2% at 800, then drops to 84.1% at 1000 and 80.6% at 1200, lower than either of the two bands below it, before climbing again through 83.7% at 1400, 86.2% at 1500, 88.4% at 1600, 91.9% at 1800, 94.1% at 2000 and 95.4% at 2200. A rating-blind report would call a missed Qxf7# at 1200 the same kind of error as a missed one at 600. Our own data says the 1200 band was, if anything, slightly less likely to deliver it than the 800 band was.

The third dataset shows findability is opening-dependent, not just rating-dependent, which a single averaged number hides completely. At the 1000 rating band, the Ruy Lopez sits at 62.7% mean findability while the Queen's Gambit Accepted sits at 10.7%, more than fifty percentage points apart at the identical rating. Rating doesn't move every opening the same direction either:

The Ruy Lopez is the single most findable opening at 1000 and actually drops by 2200, overtaken by the Sicilian Defence, Open, which more than doubles across the same range. The London System barely moves at all, gaining under two points across twelve hundred rating points of climbing strength. A rating number alone predicts almost none of that. What it's opening against, and how sharp or quiet that opening is at that specific level, predicts most of it. See what your chess rating actually predicts and the fuller ranking in openings ranked by findability for the rest of the fourteen we tested.

Which one do you actually want

If the question is "what's the strongest move here," you want a classical search engine, Stockfish, free and unlimited on Lichess's own analysis board. If the question is "would a player at my level have found it," no engine in the first two categories answers that at all; you want a human-move model like Maia, which is also the harder thing to find packaged as an opponent rather than a research repository, a gap covered in how to actually choose an opponent. If you want something to talk to about chess in plain language, an LLM narrating a real engine's output is a reasonable narrator; an LLM generating chess judgment from scratch, including basic legality, still isn't reliable, and the gap hasn't closed as the models got bigger. If you want structured, all-in-one coaching and don't mind pricing that takes checking directly, Chess.com's Diamond tier does more of that than anything else in this piece. And if what you actually want is a human-move model as the opponent and an engine for the review, in one loop instead of three separate accounts, that's the narrow thing Chessdrive is built for. That is a packaging claim, not a technology one. The technology itself is free, scattered across several other people's projects, and this piece just told you where each piece lives.