But why do I need a coach when computers know everything?

The spaces in which humans excel compared to machines seems to be contracting by the day. It is something not lost on - and often embraced by - the chess community. There was a time when it was hard to conceive of computers surpassing our species' best talents when it came to just playing chess. For a long time, positional evaluation was so beyond the silicon mind that long-term strategy and positional nuance gave our Grandmasters an edge that made computers not much of a challenge. But processing power improved, we learned more about programming, and eventually a computer beat world champion Garry Kasparov in a 6-game match. While many may have seen this as the beginning of the end for human dominance of the game, there was still a debate to be had. Kasparov had lost for very human but not very "chess" reasons: he resigned a drawn position in Game 2, after suspecting human interference, and never recovered psychologically. Still, Kasparov had won the first game convincingly, and the outcome of the match was not settled before Game 6, even if the eventual 2-4 score line was decisive. However, when Hydra thumped Michael Adams 5.5-0.5 in 2005, the writing was on the wall. This was evidence that computers would soon be the best chess players, if they weren't already. Nowadays, the matter is beyond doubt. Our best humans will never be as good as the machines.

But it is one thing to be good at the game; it is a different skillset that is required to coach someone else. Maybe not something else - Google has proved, through AlphaZero, that a machine can teach itself to play chess well. Chat GPT has made similar, albeit slower advances. But how to teach humans to get better at the game? To some extent, computers can do this - even if they are programmed and otherwise mining information provided by humans. They are certainly a useful tool - who among us, if we are serious about improving our game, doesn't use computers to analyze games or check preparation? We couldn't do without them. It's an open question (and not one that I intend to answer) as to whether AI will eventually advance to a point where it can mentor a human better than a human can. But as of right now? My hot take is that the best human coaches are far better teachers than computers can be. Anyone that wishes to improve significantly ought to invest in an actual coach, assuming they have the resources to do so. I'm not saying one can't learn from computers alone - I don't have a coach and have steadily improved over the last few years. But a human coach can take you further and get you there faster. I don't dislike the level I'm at now, but I'm sure that if I had the time and resources available to invest in a coach, I could have been even better.

How can this be? Don't computers know everything? Um... no. If you analyze a game with a computer, it can tell you what the best move was on each move, and whether the player(s) found it or not. The computer can evaluate the move that was chosen, compared to the "best" move available. It may award exclamation marks, question marks, or a combination thereof, accordingly. But it absolutely sucks at explaining why. Oh sure, it can tell you things like: "Moving your knight there forks two pieces" or "this creates a passed pawn". Thanks, Skynet. Would never have figured THAT out myself. Why is forking those two pieces better than the skewer I had available? Why is creating a passed pawn in this position a bigger deal than recovering material - and how do I calculate these outcomes when they are not consistent? Observe:

This is a position one of my students sent me. He had the Black pieces, and was on move. Things were looking good - Black had just won a Queen for knight and bishop. But my student didn't know what to do next, made the wrong choice, and went on to lose the game. Subsequently, the computer told him that 24...e4 was the right move in this position. He wanted to know why, and I sent one of my customary epic emails going into detail with the answer. Good luck getting that kind of dedication from a computer! But hang on - it gets more complicated.

My student's choice was Rf8. This, I explained (hopefully constructively) was a poor choice. The attack on the f5 pawn is superficial, since White can easily defend with 25.Be4 (which is what happened), and this creates a "pawn" chain along the c2-f5 diagonal, which locks the whole position down. Black has no real breaks in the position to open the game up, thus the Queen's scope is limited. White's extra piece on the board is more telling than the quality of pieces held, because the Black Queen cannot dominate vast swathes of the board as she would like. White went on to launch a brutal Kingside attack; Black's pieces got tangled, and White emerged victorious. Had White been on move in the position in the picture, White would probably have played Be4. So Rf8 forced a move that White wanted to make anyway - not usually a good idea. I'm yet to see a computer give this type of explanation for why NOT to make a move. The ability to explain why certain moves don't work when others do is something we mortals still have the edge on. How many times have you seen a blunder explanation from a computer say something simplistic like: "this loses a piece"?

Even with this explanation against Rf8 in hand, it could serve to make 24...e4 more confusing. After all, Bxe4 is still a distinct possibility, which gives the impression that Black is handing White that chain again on a silver platter, with a pawn into the bargain. Show me the computer that can satisfactorily explain, to someone learning the game, why allowing one particular positional structure is bad, but allowing the same structure minus one pawn is good. Other than showing you the follow-up, the computer can't do that.

Now, I could launch into a detailed explanation about good versus bad bishops, piece scope, open v closed positions, King safety and deflection, but I've already written it out once, so if you're interested, ask me for a copy of the email. But since I've left the question hanging: if 25.Bxe4, Rxe4!! is the follow-up. This cedes White even more material, practically giving away any material advantage Black had, but it increases Black's overall advantage. 26.dxe4 c4 and Black is ripping the position open so the Queen can run riot. White still has a significant army, but it is uncoordinated and disjointed. Still, I would venture that if you're not already sufficiently advanced to know these concepts, the nuances of the position are unlikely to reveal themselves to you when the computer coldly shows you the moves and maybe offers a few words for explanation here and there.

Even if the computer keeps the explanation simple, or at least the explanation is something you understand, that may still not be helpful. The reason the computer gives you as to why you should or shouldn't make a move may not be accurate. I've lost count of the amount of times I've seen a computer refer to something that isn't the most relevant thing in a move explanation. Of course, I hadn't been actively storing examples, so I can only offer the following. Exhibit A:

I was Black here, and my ego insists I point out that (a) this was my ritual "play against the computer to see if I'm awake yet" game (if I lose, I go back to sleep) and (b) I ended up having to get up, because I went on to win this game. However, the computer's explanation is not particularly helpful. Yes, OK, my move was a mistake because the b3 pawn was hanging. Tired Alisdair missed that - or did he? If memory serves, the reason I chose Qa3 was because I was worried about the Queen getting trapped if I didn't move it; that Rxb3 might lead to some complications where my Queen is forced away and then my rook hangs, or at least I have to swap off my Queen for two rooks - and I wasn't convinced that trade was better for me in the position. That's the questions I had about the position - what the fudge does the fact that I've created a passed pawn have to do with it? That's helpful, yes - but it is a bonus reason for playing Rxb3, not the main idea, and not something that will help me choose Rxb3 since it wasn't what concerned me. Obviously I value creating a passed pawn for myself, but I had bigger fish to fry at the time.

Exhibit B:

I was a little more awake here, as I was White, and dominating. But what is with this explanation for Black's move here? Oh whoopdee-doo, Black is developing a piece! That is nothing to do with the rationale for Rh4 - the rook was under attack! But let's put my coach hat on and imagine that it is one of my students that has played Rh4. I'm not much of a coach if my analysis mentions the development of their rook! The rook MUST move, because White is threatening Nxh8. So I'd be saying to my student: "OK, I see why you moved the rook to h4, because it is the only "safe" square. But if you look a move further, you can see that White can always trap the rook anyway - here they can play 23.g3 for instance. If you realise this ahead of time, you can consider other possibilities, and you might find 22...Rg8. White CAN take your rook here, but gives up the e6 knight for doing so (and in fact the best move is 23.Qf5)".

At this point, one might think: "OK, so the computer maybe sucks at explaining why you should make a certain move, but at least it tells you the best move, right? RIGHT?" Um... about that. So if you have Stockfish (other programs are available) running at max depth and give it as much time as it wants to think, it'll come up with a definitive answer on what the "best" move is. But that's not how most of us analyse games. We run a quick analysis on chess.com with a limited depth engine connecting to servers on a variable bandwidth - there's a number of things that could affect the machine's output. In simple terms: analyse the same game with allegedly the same computer program twice, and it won't analyse it the same way. I've had the odd game where I've managed to avoid any blunders, mistakes, or inaccuracies in an initial analysis, but when I look back at that game later and the original analysis isn't attached any more, the engine will look again and find errors. Case in point:

I've had this opening, and this exact position, many times against the computer. As you can see, today it decided that the move 11.Bc4 was an inaccuracy. And I mean today. The reason I played 11.Bc4 in the first place is that the computer has, on more than one occasion, told me that 11.Bc4 is the best move. I believed it - but the computer has apparently changed its mind, and this is not an isolated incident. I am constantly looking at lines in my opening preparation, wondering why this line merited being noted down when the best engine moves weren't, before realising the line I've noted down most likely WAS the engine line when I was taking notes.

Speaking of opening lines... how do you study your openings with the aid of a computer? Do you play through the engine's top recommendations, and memorise them? Well, if you're capable of doing that, all the better for you - except you might get caught out by the previous issue - the computer might change its mind as to what the best line is. But if, like most of us, you're - well, only human - then this might be too much to ask in some openings. It might be better to adopt a line that is not quite as good, but has an intuitive plan or line, so that you will remember the strategy easily in the heat of battle - you might temporarily forget the moves, but if you understand the opening well, it will come back to you. For example, playing the Black side of a King's gambit. We all know we accept, right? But after 3.Nf3, g5 is the top engine move. This invites a complicated game with lots of theory, and there are several different branches White could take you down. Each branch requires its own handling so you have to spend ages learning it - which is why at least one titled player I know just plays 3...d5 - giving them a smaller edge, but one that they'll remember how to convert much more easily. Certainly I have recommended the "second best" line in some openings, because the "engine line" is complex, does not follow a set plan, and the moves are different depending on what the opponent does, making it as hard as possible to remember.

This issue of practicality is, of course, beyond the silicon's comprehension. It is programmed to churn out the best moves. As a human, playing against another human, you weigh in other factors, such as psychological impact, time control, and how comfortable you will be with the resulting position. Just this week, for instance, I made a move that relied heavily on the first two of those factors. It was objectively not a good move, and the eval bar went from my favour to my opponent's - but it looked good - in fact, it looked so good that my opponent resigned. A computer can never explain the benefits of such a move, because it would be impossible to program in how to evaluate this. Or, to finish with a more fun example:

The computer gave me a "?!" for this - and arguably objectively, rightly so. I should promote to a Queen and mate quickly. It's not the computer's job to encourage trolling. But I could also argue that this was the right move for me to make at the time. The game was won, and my opponent wasn't a human so couldn't be embarrassed. So, I took the opportunity to practice my B+N v K ending. Practicing technique was far more important than finding the quickest possible mate in a position where promoting to a Queen requires next to no skill to finish the game off.

The one slight discomfort I have with my claim is that the hypothesis is not testable, which arguably leaves the question open. It's not like we could take one human student, create an exact replica, and then have a human train one student and a computer train the other. There is obviously a place for computers and engines in the use of coaching. We all use them, and they remain an essential tool for those of us who don't have the budget to hire a coach. I'm sure that advancements will be made in this field to make our mechanical friends even better for this purpose than they are now, and it may be useless to speculate just how good they'll be. However, I hold out hope for my species that this is one area where the machines will never quite catch up with us. In the meantime, I am absolutely convinced that we humans have the edge. If you want to improve your game, and you can hire a coach, I wholeheartedly recommend that you do so.