AI Chess Coaches?

Over the past few years, we have seen an unprecedented acceleration in artificial intelligence. What began as narrow, task specific systems has rapidly evolved into large language models capable of reasoning, explaining, summarising, and interacting in ways that were simply not possible before. This shift is not just about automation or raw performance. It represents a fundamental change in how software can communicate knowledge. Systems can now engage in dialogue, adapt explanations to context, and operate in domains that require structured reasoning rather than simple pattern matching. As a result, areas traditionally dependent on human instruction are being reexamined through a very different technical lens.

Chess has not been immune to this shift. Across the industry, companies and platforms are actively exploring how modern techniques might reshape the way players learn and improve. For most chess players worldwide, access to a human coach is limited or nonexistent, constrained by cost, geography, or availability. Engines are immensely strong and databases are vast, yet neither truly solves the core learning problem. Understanding ideas, plans, and mistakes at a human level remains difficult. The promise of these tools lies not in replacing strength, but in narrowing that gap and making structured guidance available to millions of players who currently study alone with little feedback.

But this raises a more fundamental question. Is an AI chess coach actually possible? Not in the marketing sense, but in a meaningful educational one. As someone with over twenty years of experience in data and business intelligence, I naturally approach this problem by breaking it down into systems, signals, and constraints rather than aspirations. I have spent a considerable amount of time thinking about where current approaches fall short, what technical limitations exist, and what a realistic architecture might look like if the goal is to teach rather than merely analyse.

In this article, I want to share those thoughts, the ideas, the challenges, the technical directions, and the trade-offs involved in attempting to build something that resembles a genuine chess coach.

What Do We Actually Mean by AI?

A useful place to start is with an uncomfortable statement. Artificial intelligence, in the general sense, does not yet exist. Despite the language used in headlines and product marketing, the systems we interact with today are not intelligent in the human sense, nor are they close to being so. Most credible projections suggest that anything resembling general intelligence, systems that can reason autonomously across domains, form goals, and truly understand context, remains decades away. Much of what is described as AI today is, in reality, a convenient label applied to a collection of advanced but narrow techniques, amplified by marketing and public imagination.

So if AI does not exist, what do we actually have?

The most visible shift came with the public release of large language models, most notably through tools like ChatGPT. These models are trained on vast amounts of text and learn to predict the next piece of language in a sequence with remarkable accuracy. Through scale in both data and computation, this simple objective produces surprisingly rich behaviour. We see coherent language, contextual answers, explanations, and even what appears to be reasoning. To most users, the experience feels conversational and adaptive, which naturally leads to the perception that they are interacting with something intelligent.

However, this perception is largely a mirage. Large language models do not understand concepts in the way humans do, nor do they possess intent, awareness, or grounding in the real world. They operate by identifying and reproducing statistical patterns in language, guided by probabilities rather than comprehension. When they explain an idea, they are not reasoning it through from first principles. They are assembling a response that looks like how such an explanation is usually written. This limitation is also why we have the term hallucinations. A model can produce information that sounds confident and coherent, yet is factually incorrect or entirely fabricated, because nothing in its design enforces truth. It only enforces plausibility.

That does not mean these tools are not useful. While large language models are not AI in the strong sense, they are exceptionally powerful interfaces for structuring, contextualising, and communicating complex information. For the problem we are exploring, building something that behaves like an automated chess coach, this distinction matters. The goal is not to create an intelligent being, but a system that can analyse, explain, adapt, and guide learning effectively. Even if the intelligence is an illusion, the practical value may still be very real.

How Can We Build a Chess Coach?

Now that we have a clearer understanding of what people actually mean when they say AI, the next question becomes more practical. How can we use these tools, with all their limitations, to build something that resembles a chess coach?

My first attempts were, in hindsight, the most obvious ones. When I initially started thinking about using what we loosely call AI in chess, the idea was simple. Could I send a game to a model, a PGN or a FEN, and ask for feedback? Most chat based systems expose APIs that appear to reason, explain, and analyse. On the surface, this seemed promising. In practice, it fell apart very quickly. The responses were often almost correct, and that word almost turned out to be fatal. Pieces would appear that were not on the board. Checks would be described where no check existed. Advice would be given based on positions that were subtly, but critically, wrong. In chess, there is very little margin for error, and confidence combined with inaccuracy is worse than no answer at all.

Around the same time, Google hosted an AI chess championship that further reinforced this reality. Even with safeguards in place to filter out illegal or clearly incorrect moves, what emerged was play that hovered around 800 to 1,000 Elo. The commentary team themselves acknowledged this limitation. That level of play may be interesting as a demonstration, but it is nowhere near strong or reliable enough to function as a coach. This made one thing clear very quickly. A pure chat model, no matter how impressive it sounds in conversation, cannot be trusted as a standalone chess coach.

That forced a change in direction. If a language model cannot be trusted to understand chess on its own, perhaps it should not be asked to. The next idea was to shift its role entirely. Instead of asking it to analyse positions, what if it simply explained analysis produced elsewhere? This is where external sources came into play. Engines like Stockfish, Maia, and Leela Chess Zero, along with large databases such as ChessDB or the Lichess opening database, already provide extraordinarily rich and accurate information. Evaluations, move probabilities, trends, and statistics are all available in abundance. The question then became whether a language model could act as a translator rather than a thinker.

Surprisingly, this approach worked far better than expected. By setting the temperature, effectively the creative freedom of the model, to zero, the behaviour changed completely. The model was no longer encouraged to invent or speculate. Instead, it focused on turning structured data, numbers, and move lists into readable explanations. In that role, it performed well. It could summarise trends, describe candidate moves, and explain evaluations in a way that was accessible to humans. In some cases, it could even highlight patterns across large datasets in a way that felt genuinely useful.

There are also open source efforts exploring this space from a more community driven angle. One example is ChessAgine, a new open source initiative focused on combining strong chess analysis with natural language interaction in a more structured and grounded way. Rather than treating a language model as a standalone intelligence, ChessAgine is designed around the idea that models should sit on top of engines, databases, and curated chess knowledge, acting as an interface rather than an authority, an assistant layer than being an AI Coach. A key part of this approach is its use of Model Context Protocol style integration. This allows the system to pull in concrete information from external sources, such as engine evaluations and opening databases, and then discuss that information in natural language. In practice, this makes the system feel less like an oracle and more like a second set of eyes. A player can ask about a position and receive explanations of common plans, context around typical ideas in familiar structures, or insight into why certain approaches tend to work or fail in practice.

But something was still missing. Describing a position, no matter how clearly, is not the same thing as coaching. A coach does more than explain what is happening on the board. They decide what matters, what can be ignored, what a player should focus on right now, and why a mistake is important for that specific individual. That gap between explanation and instruction is where I found myself spending countless hours thinking, sketching ideas, and questioning assumptions.

Before going any further, it is worth stepping back and addressing a more fundamental question.

What Is a Chess Coach?

While my professional career has been rooted in data and business intelligence for over twenty years, I have always had parallel passions outside of technology. Before chess became central to my life, that passion was salsa dance. I spent more than a decade teaching in London, working with students of all levels and backgrounds. That experience shaped how I think about learning far more than any technical role ever could.

Through that lens, I began thinking more deeply about the role of a chess coach or teacher. To me, a true coach is someone who can teach both the science and the art of a subject. In chess, the scientific side is relatively well defined. It includes the rules of the game, how the pieces move, basic tactical patterns, fundamental endgames, and the structured ideas behind the opening phase. These elements can be taught, drilled, tested, and measured. Engines and platforms already support this side of learning extremely well.

The art of chess is much harder to capture. This is where imagination, intuition, and conceptual understanding live. It is the ability to look at a position and sense its character, to understand long term plans rather than short term tactics, and to feel when a position calls for patience, aggression, or restraint. These qualities are not easily reduced to rules or evaluations. They are developed through guidance, analogy, and experience, often shaped by how a coach frames ideas rather than the moves themselves.

This is why platforms like Chessable, while excellent for drilling opening lines, will never fully replace a coach. Memorising moves is not the same as understanding why those moves work, when they fail, or how to adapt when an opponent deviates. Making an opening come alive requires imagination and context, and that comes from the marriage of science and art rather than repetition alone.

Another way to think about this is through music. Two pianists can be given the same piece of sheet music and play every note correctly. Yet the way the piece is expressed can feel completely different. Timing, emphasis, emotion, and interpretation are what separate mechanical accuracy from meaningful performance. Chess is no different. A coach does not just teach what to play. They teach how to think, how to imagine, and how to express ideas on the board.

The Current Challenge

If the goal is clear, a fair question follows. Why has this not already been done?

For me, the answer lies in a single stubborn problem. Defining a chess position in a way that supports coaching rather than analysis.

I spent a significant amount of time on a personal project called Minerva, with the hope that it could heuristically describe a position in human terms. Starting from a FEN alone, I worked on extracting meaningful attributes. Whether a king was in check, the presence of forks, pins, open files, pawn majorities, weak squares, and structural imbalances. Over time, I was able to generate more than thirty distinct characteristics directly from the position.

Where things began to break down was at the level that actually matters to a coach. The plan.

Consider a Caro Kann exchange structure. There are well known ideas associated with this type of position, such as queenside play or a potential minority attack. These ideas can be useful reference points, but they are far from automatic and often not the best choice. Whether they make sense depends heavily on piece placement, timing, and the broader dynamics of the position. Chess is not about applying plans by rote. It is about choosing what fits the moment.

This is where the problem becomes truly difficult. Chess contains an enormous number of unique positions, each with subtle differences that can invalidate otherwise reasonable ideas. If plans are context dependent, how does a system decide which ones matter and when?

One possible direction was to look back to engines, not for answers, but for intent. By analysing Stockfish lines, it is sometimes possible to infer what the engine is preparing. If it consistently manoeuvres pieces toward certain squares, delays pawn breaks, or avoids commonly suggested plans, that behaviour carries meaning. From there, one could attempt to work backwards and determine what the engine is aiming to achieve.

The challenge is obvious. We are attempting to translate the motives of an engine operating at a level far beyond human intuition. Interpreting the intentions of a 3,700 Elo player and expressing them in human terms is no easy task. It is brittle, indirect, and full of edge cases.

Yet despite its difficulty, this may be the most promising direction. Not because engines think like humans, but because their consistency can reveal what truly matters in a position.

So What Can Be Done?

Given these constraints, the question shifts from what is ideal to what is realistically achievable. If I continue to work in this space, the next step is unlikely to be a perfect chess coach. Instead, it is something more modest. An assistant focused on teaching the basics of chess well.

At this level, heuristics and structured data are far more reliable. Hanging pieces, missed captures, undeveloped pieces, repeated violations of basic principles, and persistent structural weaknesses can all be detected with a high degree of confidence. These are not subjective judgements. They are concrete signals that point to fundamental problems in play.

By importing a player’s online games and combining this data with engine analysis and opening books, it becomes possible to highlight meaningful trends. Did the player leave theory unusually early? Are there openings where they consistently drift into worse positions by move ten? Are there recurring tactical oversights across many games? These are questions engines and databases are well equipped to answer.

This approach has limits. It will not teach imagination. It will not replace a coach when it comes to long term planning or positional judgement. It will only take a player so far. But that does not make it insignificant. For many players, building a solid foundation is already a major challenge. Learning chess is hard, and progress often stalls due to repeated uncorrected fundamentals.

A free or accessible companion that consistently points out these issues, explains why they matter, and suggests areas to study still has real value. At the very least, it feels like an interesting and worthwhile use of my time. And who knows, it might even form the backbone for something more ambitious in the future.

Closing Thoughts

I do believe that AI based chess coaches will one day be able to provide an experience that feels close to working with a real coach, at least from a technical perspective. With strong engines, quality data, well designed heuristics, and careful use of language models, the gap between tools and teaching will continue to narrow.

What I do not believe is that an AI chess coach will ever fully replicate the most important part of the learning experience. The human connection. A good coach brings empathy, intuition, and understanding that go beyond the position on the board. They read the student, adapt their approach, and respond to confidence, frustration, and motivation. That relationship cannot be automated.

For that reason, I see AI chess coaches not as replacements, but as tools. Useful ones, potentially powerful ones, but tools nonetheless. They can support individual study, handle technical groundwork, and free human coaches to focus on creativity and connection rather than repetitive analysis.

So while I am optimistic about what is possible, I remain realistic about what should be expected. I hope you found this article interesting. More than anything, I wanted to share my thoughts on a topic that sits at the intersection of technology, chess, and teaching. I would love to know what you think, so please feel free to leave a comment.

Kind Reagrds,
Toan Hoang (@HollowLeaf)