Inspired by https://www.lesswrong.com/posts/nR3DkyivzF4ve97oM/how-go-players-disempower-themselves-to-ai and how AI interacts with Go players, I thought it would be interesting to share a collection of anecdotes and experiences with chess.
I’ve purposefully left this article without any chess diagrams or positions, which may come in a Part II. In this piece I use ‘engine,’ ‘computer,’ and ‘AI’ somewhat interchangeably.
In 1997, Deep Blue defeated then reigning World Champion Garry Kasparov in a match, 3.5-2.5. It was a historic result, although not without controversy, and judging by the quality of the games, perhaps Kasparov still showed superior strategic understanding. Human-computer matches would continue over the next decade. The final human-computer match was played between Deep Fritz and Vladimir Kramnik, with the computer winning 4-2. It has to be said that even there, the machine showed few signs of superhuman ability. The World Champion inexplicably blundered checkmate in one move in Game 2, with 4 of the other 5 games being drawn, and a loss in the final game in a last ditch attempt to win the match.
Since then, computers have steadily gotten stronger. Because current engines do not play against humans, it is difficult to get an exact grasp on their rating level (instead, the 1997 and 2006 matches are often used as a bridge to peg down their Elo). In addition, the drawish nature of chess forces developers to use imbalanced opening books to differentiate the skill gap between engines, because a match from the start position would feature almost all draws. We can say that as time goes on, there are fewer positions where engines tend to ‘misplay,’ although the space of all possible positions is enormous.
Chess players today grow up in an interesting era where computer dominance is a given. Computer analysis and programs are ubiquitous; any commentary of top level games features an ‘eval bar,’ an indicator that shows exactly how well a player is doing at any given moment. The game notation often features colored commentary indicating which moves are mistakes or blunders. On the now defunct game spectating website Chessbomb.com, chatters would derogatorily mark bad games as ‘rainbow colored.’
But the share of top players that use AI has also steadily increased. In the early days, some strong GMs consistently played their pet openings, and for a while, they could get away with it, too. One of my early coaches, a player in the top 100, steadfastly used Rybka 3 even as later versions came out. He liked its feel and how it would approach certain positions.
What happened to the players that refused the AI adoption? They gradually stopped becoming top players. Today, of the top 100 players, at least 95, and probably 100, use the latest engines or play opening moves suggested by the top engines.1
The chess world is remarkably inclusionary of all ages. In speaking to any grandmaster from the 80s, there’s an immediate sign of reverence in how they spoke of the top players of the day. Patrick Wolff described analyzing with Vishy Anand as ‘being in the presence of God,’ a sentiment later shared by a young Magnus Carlsen who had already taken the world #1 ranking. Perhaps it is possible that pre-AI, humans demonstrated a superior method of understanding. They just got it.2
I think differently. The fundamental problem of getting it is this: The goal of playing chess is not to create art, not to show understanding, not to think or display strategical brilliance. The goal of playing chess is to win chess games.3 The top chess players use whatever resources they can, including copious amounts of computer use to deeply analyze every single variant and position.
But if AI is freely available to use for everyone, what is the differentiator? Why are the elite players better than the less elite players?4 Top players go into less explored lines - with a slight twist. In popular culture, people cite Magnus Carlsen ‘purposefully not playing the best moves,’ although perhaps a more apt description would be ‘purposefully not playing the best moves, and then analyzing the next 5-10 moves and memorizing the best moves.’ It is not by accident that Magnus Carlsen rarely, if ever, gets outprepared in the opening, and it is not because he ‘doesn’t use AI.’5
While the goal is still to avoid the well-trotted lines, the use of AI has exploded, rather than diminished. It is difficult to know exactly how much analysis has been done, because so much of it is done behind closed doors. But even players on the fringe of the top-100 have thousands of opening files, each with thousands of moves. The top players, who are able to hire teams to analyze for them, surely have significantly more. In the past, entire world championship matches debated single lines, and the truth was significantly harder to come by. But now knowing that surprises are coming, it becomes necessary to analyze fringe variations, to become well prepared against all possibilities. Maybe AI users get it, maybe they don’t get it, but AI non-users are not even in the arena. They don’t have the chance to get it.6
I had a curious case of ‘not getting it,’ one year ago. By sheer coincidence, I faced off against Carlsen in the Grenke Freestyle Open, a tournament with the starting position randomized, rendering opening memorization obsolete. I reached a practically lost position with white in just 6 or 7 moves. Maybe he was just much stronger than me. Perhaps his way of doing chess, reading books, understanding the classics, sheer talent, awesomeness, or whatever. I had been duped into thinking I got it all the way into the top 30 players in the world, but he really got it.
Getting it is hard. Perhaps the ugly truth is that most, if not all, people spend their whole lives on a singular thing only to not get it. In 1963, Bobby Fischer famously declared that he could give any top woman chess player knight odds (an almost insurmountable disadvantage) and still win. It was a comment so outrageous that it is universally decried as misogynistic. But what is almost certainly true is that AI today could also give him knight odds and win, at least at a fairly quick time control.
In 2024, I played Leela Knight Odds in a match. It was the first time AI demonstrated that it could beat a top-50 player with knight odds. Although I won the rapid portion (with more starting time), at a blitz time control, Leela won 8-2 (6 wins, with no losses). However, my result turned out to be an exceptional performance. Over the next months, the network improved, with Hikaru’s score of 12.5% in a double-digit game series being the best of any top player.7
It was a remarkable result, considering that a few years ago, Jordan van Foreest defeated Stockfish with a deficit of just two pawns, a much smaller handicap. But strong chess engines are designed to play against other strong engines and in lost positions will aim to prolong the game rather than look for counterchances. Leela was remarkable in the way it would consistently find resources when it looked like none were available.
But more impressive than its resourcefulness was its strategic understanding. Playing against it, even with an extra piece, felt incredibly constrictive. It would gain space, create weaknesses, create situations where even an extra piece felt like a disadvantage rather than an advantage. It would attack on one side of the board, then forget about it, and somehow connect together a chain of moves 15 moves later that made it all seem relevant.
It was artistic, gorgeous, a type of chess that was awe-inspiring and beautiful to play against. I asked GM Larry Kaufman for a copy of the weights so I could analyze regular chess with Leela to look for new ideas. His response was perhaps a bit depressing:
“The problem is that what the Leela Odds bots are trained to do is to find the best practical chances against a human GM level player playing blitz, given that Leela is something like a thousand elo stronger than that. So imagine that you were asked to train for how best to defeat 1700 rated players. You can’t just play them a million games and see what works, because you will win nearly every game no matter what you play.”
Perhaps we never really got it, ever, at all.8
The point I am trying to make is not that chess players use AI to empower their understanding of the game, or even that man and machine work in tandem to advance human knowledge.9 To the practitioner, the entire premise is flawed: it is simply about what AI can do in service of the goal.
Chess lives in an interesting space where we have all but given ourselves to AI. In the early 2010s, there was a reasonable claim that humans could provide input to the top engines.10 Today, engines are essentially an oracle that show the correct way in minutes, if not seconds. For all practical use cases, they are infallible.
Yet do top players show signs of disempowerment? Have we signed away our originality and humanity to the machines? I think the answer is still no.
The current World Champion, Gukesh D, famously did not use computers until he became a Grandmaster. Javokhir Sindarov, his next challenger, has extensive training camps without any computers.11 Of course, as noted above, one should take such training techniques with a grain of salt. Despite having non-computer use as an integral part of his training, Sindarov tore through the Candidates with scintillating play, but also brilliant preparation.
Even more to the point, competitive environments like chess are self-correcting. If there are gains to be made by straying away from AI, players eventually tend to use it less. Players use AI more because their competitors use AI more. As computers have gotten stronger, so too has preparation become deeper and more accurate.
But rather than seeing players today being captured by a sort of AI psychosis, top players are leveraging AI in incredible ways. When I look at the current crop of young elite players, they find original ideas, calculate incredibly deeply, and come with a mix of both strategic and swashbuckling styles. One commonly cited influence of computers is that players today find many more defensive resources than they used to.
The truth is that everything is important, and neglecting any part of the game will always rear its ugly head. Learning from AI is beyond critical, but so is training without it; mental stamina, physical fitness, sleep - all of it is critical to delivering results. Perhaps what makes chess interesting is the clash of unique styles - which stem from varying training methods - each leveraging their own advantages and disadvantages.
Chess players have disempowered themselves to AI. But with this disempowerment comes empowerment to new ideas, the creation (and destruction!) of openings, strategies, and styles. Even so, I do not think most chess players concern themselves with the nobility bequeathed upon them as guardians of the Royal Game.12
The discourse on cheating in chess has exploded in the past few years, and it is hard to avoid. One interesting phenomenon is that as chess players grow up with engines, some players who have engaged in AI misuse at a younger age have since grown up to become very strong players.
The chess community, on the whole, has decided to view online cheating as less important than over the board (OTB) cheating. Perhaps there is an element of premeditation, or preparation, that comes with OTB cheating. Chess websites are self-policing. While they collaborate with FIDE, the governing body of chess, they are not under its jurisdiction. But whatever the reason, the amount of online cheating is so prevalent compared to OTB that it perhaps makes sense to distinguish between the two.
In particular, it has come to the point where certain players who have been quietly removed from online servers have made it to the very top of the chess world. In 2024, the Candidates Tournament (the precursor to the World Championship) featured for the first time a player banned for FairPlay Violations online.13 Now in 2026, a player previously banned for FairPlay Violations will play in the World Championship for the first time.14
Perhaps to the chagrin of a few, cheating and AI misuse does not seem to materially impact the career or development of top chess players. Although it is an ethical debate and perhaps outside the scope of this article, it harkens back to the central principle. Chess players are not rewarded for honesty, integrity, or artistic value. They are rewarded for winning chess games, and AI misuse, after serving a small sentence, does not seem to hinder winning in the long term.
I am not really sure if this principle extends to the broader world.
One interesting facet of the introduction of chess engines into chess has been the rapid ascent of prodigies. Recently, Turkish prodigy Yagiz Erdogmus broke into the esteemed 2700 elo club at the age of 14, breaking the previous record by more than a year. In the 1980s, information was scarce, with chess games locked behind The Informant, a magazine which contained all the top level games from the past 6 months. Today, information is readily available, and younger players can gain experience against top level players online, and they can analyze at the same capacity as a top player, with the same tools. At just 14 years of age, Erdogmus came within a whisker of holding Magnus Carlsen to a draw with black, a sensational accomplishment.15
Thinking that AI is not useful or will not be able to do incredibly valuable tasks feels to me more a reflection of the type of tasks someone is doing rather than an indictment of AI.16 In the same way that Large Language Models are just ‘next token predictors,’ so too are chess engines like Stockfish just ‘brute force tree search algorithms.’
In 1913, G.H. Hardy was the first mathematician to respond and recognize Ramanujan’s brilliance. The mathematical community was far more insular; it was not enough to do mathematics, it was also a shared language and community. But this means that there is a barrier to entry, one which is necessary to remove the cranks, but also lets talent fall to the wayside.
In the AI era, with the proliferation of information, we will probably see even more talent and incredible things being done at a young age. Chess, while a narrow and well-defined discipline, already sees teenagers competing with the world’s very best. Perhaps the most optimistic note of today is that young talents with AI explode onto the scene and solve problems, create art, and invent architectures.