The short answer
Online mahjong on an established platform is fair in the way that counts: the wall is shuffled by the server before the hand starts, and no ordinary account can see it, predict it, or reorder it once play begins. The real risks are human rather than algorithmic — two friends sharing hands over a voice call, an unlabelled bot grinding a ranked ladder, or a player reading moves off an engine mid-hand.
The suspicion that a platform deals against you personally rarely survives contact with arithmetic. A digital table finishes hands several times faster than a physical one, so a losing run that would take a weekly club months arrives in a single Saturday session.
“Is it fair?” is really three separate questions
Players tend to fold very different worries into one word. Pulling them apart is the quickest way to judge whether a platform deserves your time, your ranking, or your money.
| Fairness claim | What it means | What can go wrong | How you check it |
|---|---|---|---|
| Shuffle integrity | The tile order is random and fixed before the deal | Weak randomness, or a wall edited mid-hand | Downloadable replays; an audited generator for money games |
| Opponent integrity | The other three seats are who they claim to be | Undisclosed bots, colluding pairs, borrowed accounts | Labelled AI seats, published ladder rules, a working report button |
| Information integrity | Nobody receives data their seat should not have | Client-side leaks, spectator exploits, live assistance software | Server-authoritative design, delayed spectating, anti-assist policy |
Almost every accusation of rigging points at the first row. Almost all genuine cheating happens in the second and third.
How a digital wall is actually built
A mahjong server keeps the tile set as an ordinary array — 136 objects for Japanese riichi, 144 for most Chinese and Hong Kong rule sets, one reason the two families feel so different once you look at how their scoring systems work. It then shuffles that array with a Fisher-Yates pass: as Wikipedia’s article on the Fisher-Yates shuffle sets out, you walk the list once and swap each position with a randomly chosen earlier one, which produces every possible ordering with equal probability.
The randomness comes from a generator seeded before the deal. Casual apps use a general-purpose pseudorandom generator; software handling real money is expected to use a cryptographically secure pseudorandom number generator, a class of generator defined by the property that a long run of past outputs still tells an attacker nothing about the next.
The structural fact matters more than the algorithm. On a competently built server the entire wall exists before your first draw, so there is no mechanism that notices you are winning and slips you a dead tile. Platforms that store full replays have also committed to a wall they cannot quietly revise.
What certification does and does not prove
Independent testing houses (Gaming Laboratories International, eCOGRA and iTech Labs are the names that recur in licensed gambling) examine random number generators against a licence. A free mahjong app has no such obligation and rarely seeks one.
Missing certification is not evidence of a crooked shuffle; it means nobody outside the company has inspected the code. Be relaxed about a casual ladder, and careful about anything that takes a deposit.
Why a fair deal still feels rigged
Take a perfectly random wall and four evenly matched players. Each wins roughly a quarter of hands, so four consecutive losses arrive about 32 per cent of the time, and a ten-hand drought turns up in roughly one stretch out of eighteen. Neither is a glitch.
Then add hand rate. Tiles shuffle themselves, scoring is instant, nobody hunts for a lost stick, and turns time out in seconds. An online session packs in several times the hands of an evening around a table, compressing a year of a monthly club into an afternoon.
Memory finishes the job. Nobody logs the hand where they reached tenpai (one tile away from a complete hand) on turn six; everyone remembers the third deal-in of the night, meaning the third time they discarded the exact tile an opponent was waiting on. Judge a run by your placement average over a few hundred games, and check whether the pattern is variance or one of the habits that quietly cost beginners hands.
AI opponents: labelled is fine, hidden is not
Bots fill empty seats when matchmaking is thin, run tutorial modes, and give a beginner something to practise against at two in the morning. Used that way they are a feature, and serious platforms mark them at the table rather than in a policy document.
Mahjong AI is genuinely strong. Microsoft Research Asia’s Suphx — described in “Suphx: Mastering Mahjong with Deep Reinforcement Learning”, posted to arXiv in March 2020 — reached 10 dan on the Japanese platform Tenhou, and open-source riichi engines are now standard review tools, part of the same shift that streaming and esports brought to the game.
The problem is never a strong bot in a declared seat. It is an undeclared bot farming ranked points, or a human piping the live game state into an engine and playing its output, which is the mahjong equivalent of engine assistance in online chess.
Which cheating vectors are worth worrying about
| Vector | How it works | Realistic risk | Usual defence |
|---|---|---|---|
| Collusion | Two players in a private call share hands and steer discards | Highest impact, hardest to detect | Randomised seating, pair-history analysis, replay review |
| Undisclosed bots | Scripted accounts grind ranked points or in-game currency | Common on free ladders | Timing analysis, behavioural fingerprinting, player reports |
| Account sharing | A stronger player pilots somebody else’s ranked account | Moderate, mostly a ladder-integrity problem | Device and session fingerprinting |
| Live assistance | An engine advises a human during the hand | Growing, difficult to prove | Explicit anti-assist rules, review of decision quality |
| Client exploits | A modified client reads state the server should never send | Rare where the server holds hidden data | Server-authoritative design; the client sees only its own seat |
Collusion leads the list because of the size of what it steals. One partner’s thirteen tiles convert a hidden opponent into a known one (the largest single information gain available at a mahjong table) and the pair can then feed each other winning tiles while playing safe against the strangers.
Note that these defences are statistical rather than preventative. Cheats get caught because results between two accounts are too good for coincidence, which is why a report with a replay attached is worth far more to a moderation team than a paragraph of frustration.
Five checks before you trust a platform
- Can you replay a finished hand tile by tile? Stored hand histories are the most useful fairness feature a platform can ship.
- Are the rules and the full scoring table published? A platform that hides its yaku list (the set of scoring patterns a hand must contain before it may legally win) is hiding worse.
- Are AI seats labelled at the table, in the moment?
- Does ranked play state a position on assistance software? Silence usually means nobody enforces anything.
- If money moves, is there a named licence and a named testing house?
The one shuffle nobody has to take on trust
Every question above exists because you cannot watch the shuffle. Around a physical table you can: four pairs of hands turn the tiles face down, the wall is built in plain view, and dice decide where it breaks.
That transparency is a quiet reason in-person mahjong has survived the convenience of apps instead of being replaced by them. The reasons digital mahjong grew so quickly are real, but a home table settles the fairness argument by construction.
Sydney players who want that certainty and do not own 144 tiles can borrow ours. Tile sets go out from A$20, tables from A$40 and the full table, stool and tile package from A$70, collected at Brodie Spark Drive in Wolli Creek or delivered anywhere in Greater Sydney, cleaned before each hire. Check what a set costs for your date and build a wall you watched go up yourself.
Fairness questions players actually ask
Is online mahjong rigged to make you lose after a winning streak?
No. On a properly built server the whole wall is generated before the hand begins, so there is no point at which the software could react to your recent results. Losing runs after good runs are what random dealing looks like at a high hand rate.
Do free mahjong apps use certified random number generators?
Usually not, because certification is a regulatory requirement attached to gambling licences rather than to casual games. That does not make their shuffles bad, only unverified by anyone outside the company.
Can AI opponents see my hand?
A properly built bot receives exactly what a human in that seat receives: its own tiles, the discard pool, the melds other players have claimed face up, and the dora indicator, the flipped tile that designates the round’s bonus tiles. A bot that could see hidden tiles would be evidence of a broken server, not clever AI.
How do platforms detect collusion?
Mostly by statistics. Systems look for accounts that repeatedly meet at the same table, lopsided deal-in patterns between specific pairs (one account discarding the winning tile to the same partner far more often than chance allows) and win rates that outlast variance, then confirm by reviewing stored replays.
Is it cheating to use a tile efficiency tool while playing?
During a live ranked hand, almost always yes — real-time assistance is banned wherever the ladder is taken seriously. Running the same engine over your replays afterwards is not cheating at all; it is how strong players study.
Hosting soon? Rent a complete mahjong setup delivered across Sydney, or browse our original mahjong merchandise.
More in Lifestyle & Social, Mahjong Variations