AI chess coach: what it can and cannot do
Software can measure your games, count where the mistakes fall, name the line you left and rebuild your own positions as drills. It cannot see why you played the move, watch you think, or promise you rating points. The useful question is not whether it is intelligent - it is which of the two lists a given claim belongs on.
What the phrase covers
An AI chess coach usually means one of three quite different things: a chess engine with a friendly interface, a chat model that talks about chess, or a measurement system that reads your games and keeps a record. They fail in different ways. An engine is right about positions and silent about you. A chat model is fluent about you and can be confidently wrong about the position in front of it.
Mikhail is the third kind. There is a chess engine underneath, a fixed set of named weakness categories above it, and a text layer that arranges facts those two produced. It is not a play server and not a chat window, and it never composes chess content of its own - when the backend has nothing grounded to say, the page says less rather than filling the space.
What software can genuinely do
It can check every move of every game you played, which no human coach has the hours for, and keep the ones where the loss was real. It can count: how often each kind of mistake happens, which phase it happens in, which opening families it clusters in. Counting is the part people cannot do about themselves, because memory keeps the painful games rather than the frequent ones.
It can also do bookkeeping that is dull and reliable. Tracing your moves through a catalog of named lines and reporting the exact ply your game left the book. Turning a position you misplayed into a drill with a gradeable answer. Holding a mastery number that rises only on a clean, unassisted solve, so it cannot be talked upward. None of that is insight, and all of it is work you would otherwise not do.
What it cannot do, structurally
It cannot see why you played the move. The clock, the plan you settled on four moves earlier, the fact that you had already decided your opponent was better - none of that is in a PGN. This is not a gap waiting for a better model; it is missing data. In Mikhail one category exists precisely to record that limit: time and choice discipline is defined, marked unsupported, and can never be diagnosed, because nothing in the pipeline observes the clock.
It cannot watch you think, which is most of what a good human coach does - noticing that you stopped considering your opponent's ideas, or that you always take the first plausible move in an unclear position. And it cannot promise rating points. No such promise is made anywhere on this site, because the honest version would require knowing how much you will play, against whom, and what you will do with what you are told.
The failure mode to watch for
The characteristic failure of coaching software is not being wrong. It is always having an answer. A system that must fill a "your main weakness" slot will fill it from whatever it has, and thin evidence produces a confident sentence rather than a blank space. That sentence is then trained against for a month.
The defences are unglamorous and checkable. Evidence that cannot be attributed to a named category is filed as unclassified, and unclassified can never win a top weakness. Categories that name a consequence rather than a decision were retired as diagnoses instead of being kept because they sounded insightful. The coach report stays silent below two engine-reviewed games, and statistics sections that lack the data behind them say so instead of drawing a line through three points.
How to use one without being managed by it
Read the counts, not the adjectives. Ask what evidence is behind a claim and how much of it there is; a claim that comes with a game and a move number can be checked in thirty seconds, and one that does not is an opinion with a computer's voice. If two tools disagree about your biggest weakness, the disagreement is telling you that neither has enough evidence yet.
Then do the part that is unavoidably yours. The software can hand you the two positions where your game changed; the reconstruction of what you believed at the time is the work that changes the next game, and it takes ten minutes and a pen. A coach who measures well is a genuine advantage over having no measurement at all. It is not the same thing as someone who knows you.
Questions
Is Mikhail a chatbot?
No. There is no conversation surface: you get engine reviews, counted weaknesses, drills, an opening database and a statistics page. The prose you read is assembled from facts the analysis produced, which is why it is sometimes shorter than you would like and never invents a chess claim.
Can it replace a human coach?
It replaces the part of coaching that is measurement and bookkeeping, and that part is large. It cannot replace someone watching how you decide, asking why you rejected a move, or noticing that your problem this month is confidence rather than calculation.
Is it really free?
Yes - there is no paid tier and no feature held back behind one. An account is needed for the parts that store your games and your counts, because they have to live somewhere, and the opening database is readable without one.
What if it says nothing about my play?
Then it has nothing measured to say, which is the intended behaviour rather than a fault. Play and review more games; the report starts once there are counts behind it, and an ordering invented from four events would tell you less than the silence does.
Does it use my games to train a model?
Your games are analysed to produce your own review, your counted weaknesses and your drills. Game history is written to disk only after you turn game storage on, and turning it off stops it - the data page on /about states exactly what is kept.
Try it on your games
The parts of Mikhail this guide describes are Coach report and Statistics, which need a free account — your games and your counts have to live somewhere.