AI transparency

1. You are talking to a machine

The AI Coach is a large language model. It is not a chess teacher, not a ChessBase author and not a person, and it can be confidently wrong.

Everything it says about a position is generated on the spot — or read back from a cache of earlier answers — and none of it has been reviewed by a human before you see it. The narration you can play over an article is a synthetic voice reading generated text. Treat both as a well-read opponent talking you through a position, not as a reference.

2. Which models, and whose account

The Coach is bring your own key: it runs on models from Anthropic, OpenAI and ElevenLabs (the last one for speech), called from your browser with a key you supply, on your own account with that provider. Which model answered a given article is written in its byline.

Without a key you get the cached answers other readers have already paid for, which is why the Coach works at all for a visitor who has never entered one. See the privacy notice for where your key and your questions go.

3. What on this site is AI-generated

• Coach answers, whether generated for you or served from the cache.
• The Article of the day on the home page. Its byline names the model that wrote it and the date, and says so in every language it is offered in.
• Editorial prose the Coach contributed, wherever it appears in the app.
• The narration audio. This is synthetic speech from a synthetic voice — a voice built by a text-to-speech provider, not a recording of anyone and not an imitation of any real person.
• Translations. Every language other than English is a machine translation of the English answer, checked against a reviewed glossary of that language’s established chess vocabulary.
• The Coach’s avatars. The chess pieces with faces are generated geometry, ours, and deliberately not human likenesses — no likeness rights and no deepfake question to answer. The one human figure on offer is a public-domain model of nobody in particular.

4. What is measured, not generated

This is the more important half. Most of the numbers in Theory are mined from master games and carry their own provenance — how many games, which rating band, which years — so you can check them:

• LiveBook statistics and game counts come from the ChessBase databases. They are counts.
• The opening character profile (sharpness, decisiveness, closedness, breadth, endgame tendency) is computed from a sample of master games through that position, evaluated position by position. No model was asked what it thought of the Najdorf.
• The recurring plans under “Key ideas” are the moves master practice actually converges on past that position, with the support count, the number of distinct lines they recur in, and example games you can open.
• The opening catalog itself, its names and its ECO codes, is ChessBase reference data.

The reason for the split is simple: a mined number has provenance and a reader can check it, while a label a model attached to an opening is an opinion nobody can check.

5. How a published answer is checked

An answer that gets cached is shown to other people, so it goes through a machine check before it is shared — not a review of its opinions, but of its chess:

• Every line replays. Each variation an answer links to, and every line a narration plays on the board, must be a legal sequence of moves from the position it claims to start at.
• Every line survives an engine walk. No move in it may throw away roughly a pawn and a half or more. A line that does is not shared, whatever the prose around it says.
• One verdict per position, in every language. Only the English answer is analysed; the others are translations of it at the same position. Two languages therefore cannot publish different conclusions about the same move.

The check fails closed: if it cannot verify a line, the answer is simply not published. It says nothing about whether the assessment is any good — only that the moves are real and not blunders.

6. What we do not do

• No automated decision-making about people, and no profiling. “Find your opening” answers five sliders you set yourself, with arithmetic you can see the inputs of; it keeps nothing about you.
• No emotion recognition and no biometric categorisation.
• No deepfakes: no synthetic image, video or voice of any real person.
• No camera, no microphone, no AI-driven advertising.
• Nothing you type is used to train a model by us. What your own key sends to your own provider is between you and them.

7. Found a mistake?

Please tell us. A cached answer is visible to everyone who reaches that position, so a wrong one is worth more than a shrug — and a wrong line is exactly the kind of error the machine check is not able to catch. The ChessBase contact page reaches us.