Mahjong AI Opponents: Bots That Don't Cheat
If you've played mahjong against a bot and felt it was unfair, like the bot somehow always had the right tile, or always knew what to discard, you weren't imagining things. A lot of mahjong AI either cheats outright (peeks at your hand) or plays at random (no strategy at all). Neither feels like a real opponent.
This is the page that explains how mahjong AI is actually built, why some bots feel cheap, and what to look for in a good one.
Three kinds of mahjong AI
Most mahjong games use one of three approaches:
- Random / weak rule-based. The bot picks tiles to keep based on simple heuristics (keep matching tiles, discard isolated ones) without considering what hand it's building. Easy to beat but feels boring.
- Cheating rule-based. The bot can "see" your hand and the wall. It plays optimally because it knows everything. Feels infuriating, especially when it discards the exact tile you needed.
- Strategic AI (rare). The bot only sees what a human player would see (the wall, discards, exposures) and uses real strategy: hand selection, the Charleston, defensive play, opponent reading. Hard to build, but feels like a real opponent.
Most browser-based and free mobile mahjong games use option 1 or 2. Option 3 is what serious mahjong players actually want.
How americanmahjongg.app's AI works
Two difficulty tiers:
- Sparrow (free, instant). A logic-based bot that scores tiles based on hand-completion potential and discards strategically. Doesn't cheat, only sees what a human player sees. Predictable enough that experienced players will out-play it, gentle enough for new players to win.
- Dragon (free to you, costs us a few cents per game). A Claude-powered AI that uses pattern recognition and strategy across the full NMJL card. Reads opponents' exposures, considers which hands are still viable, manages Jokers strategically, and adapts mid-game. Costs us about $0.02 per game in API calls. Doesn't see your hand. Doesn't see the wall. Just plays mahjong.
The Dragon mode is particularly useful for practice. It plays well enough that beating it actually means something. It also doesn't make the same mistake twice, you can use it to identify weaknesses in your own play.
Why this matters for learning
Mahjong is an information game. Watching how a strong opponent plays, what they keep, what they discard, when they call, how they react to your moves, is one of the fastest ways to improve. A bot that cheats teaches you nothing useful. A bot that plays randomly teaches you nothing useful. A bot that plays real strategy teaches you a lot.
If you're trying to get better at the NMJL card or develop your defensive play, a strong AI opponent is the closest you'll get to having a patient mahjong tutor available 24/7.
What weak AI looks like in practice
Tells that you're playing against a weak or cheating AI:
- The bot always has the perfect discard for the situation
- The bot never seems to make mistakes or take risks
- The bot wins or loses the same way every game (no adaptation)
- The bot calls tiles you wouldn't expect, hinting it can see your hand
- The bot's strategy is suspiciously simple (always keeps doubles, discards singles) regardless of the card
Try the Dragon mode
The fastest way to evaluate any AI is to play against it. Start a game on americanmahjongg.app, choose Dragon mode in the lobby, and play through a few hands. The difference from rule-based bots is usually obvious within the first game.
Related guides
- American Mahjong strategy: how good players think
- Reading opponents: what their tiles tell you
- Online sites compared: full platform comparison
- Best free American Mahjong online