Tools and method

When mahjong AI says you blundered, is it right?

AI review is convenient, but its conclusion is conditional. Treat it as a referee and you learn the wrong lesson; treat it as a question generator and it becomes useful.

Published 2026-10-217 min readMahjong Study Club editorial team

Author · Mahjong Study Club Editorial Team · Editorial team and review process

Key points

AI advice depends on the rules, the table situation and the opponent model.
AI does not carry your risk preference, fatigue or tournament goal.
Do not chase a 100% match rate; study the alternatives and the gap.
Group AI “mistakes” into types before turning them into training.

1. AI gives a conditional answer

Most hand-review tools calculate expected value under a fixed ruleset, a particular opponent model and a particular objective. They assume you know all visible information and that opponents behave like the model.

A real table is different. Opponents deviate, the score may demand survival, and a tournament may care about one placement rather than long-run EV. Change the conditions and the ranking of moves can reverse.

2. Blind spot one: rules and table situation

The same hand has a different risk budget in an East game and a hanchan, at flat scores and in the final round, when chasing first and when protecting third. An AI calculating flat EV may give a “long-run correct, right-now wrong” recommendation.

Before accepting a verdict, ask: what were the rules and the goal of this hand? If the AI did not know, its “mistake” is only a reference.

3. Blind spot two: the opponent model

AI assumes opponents riichi, call and fold at certain frequencies. Your actual opponents may be more aggressive or more passive. If the model is wrong, the danger assessment is wrong too.

When AI says “push”, check whether it assumes the opponent will fold. If they never fold, your attack may be worth much less than the model thinks.

4. Blind spot three: result bias

AI evaluates decisions by expected value, not by outcome. You followed the AI, dealt in, and still made a good decision. You ignored it, won the hand, and still made a bad one.

Write decision quality and hand result in separate columns. Otherwise you train the dangerous instinct that a good result means a good decision.

5. Use AI to ask better questions

Do not only read the top recommendation. Read the second and third candidates and the size of the gap. A tiny gap is a style choice; a large gap is worth investigating.

Sort the flagged mistakes into four buckets: missed information, score situation, opponent model, execution. The first two are process fixes, the third is reading practice, the fourth is training and fatigue management.

Keep your own judgement. AI is a review instrument, not a coach who plays the hand for you.

FAQ

Is AI review better than manual review?

Use AI to find candidates, then explain “why” yourself. AI is good at calculation; humans are better at table situation, opponents and psychology. Combine them.

Is a higher AI match rate better?

Not necessarily. A high match rate can mean conservative play, or that you are fitting the AI’s model. Look at error types and EV gaps instead.

The AI says 55 vs 45 — should I change?

First check whether the gap is stable. A small gap may be model noise or a style choice. Validate with more hands before changing.

This is educational training material. We oppose gambling. Hands and figures are teaching examples, not income claims.

← Back to the blog