Bot detection on ClubGG
Detection on ClubGG operates across two distinct layers. NSUS runs network-wide statistical and behavioural signals across the entire population, while club owners and their agents apply human judgment with direct seat-level control inside each club. The layers have different evidence, enforcement powers, and commercial incentives. Many cheats persist in the gap between them.
Updated 28 May 2026, 14 min read
Summary
- Enforcement is distributed across two parties with different incentives: NSUS runs network-wide statistical signals; club owners and agents run human judgement with direct seat control. Catching a cheat and wanting to act on it are not the same thing.
- The club owner is the most powerful detector and the least reliable one. He sees every transfer and seat, but he also profits from the rake the cheat generates and may have vouched for the player himself.
- Most real-time bot catching at small and mid stakes is a human decision by an owner, not an algorithm verdict. The network layer mostly matters at scale and against farms.
- Outside-channel collusion through WeChat, KakaoTalk, Telegram or Discord is the dominant cheat because it falls in nobody's clear jurisdiction: the platform cannot see the chat, and the owner often has no reason to look.
- The network layer uses adversarial classification (Dalvi 2004, Lowd and Meek 2005); the club layer is a social problem. An account stays healthy through human-like timing and sizing on the first and sensible table discipline on the second.
Who is responsible for stopping each cheat?
Sort ClubGG cheating by who is supposed to stop it and whether that party has a reason to act, not by how it works. Some cheats threaten the operator or the club owner. The dangerous ones threaten only the players while leaving everyone above them better off. The table sorts the categories by who holds the primary responsibility and how aligned their incentives are with catching it.
| Category | Operator priority | Primary detector | Difficulty |
|---|---|---|---|
| Outside-channel collusion | Highest | Statistical showdown correlation + club owner | High |
| Chip dumping inside a club | High | Account graph + transfer logs | Low–Medium |
| Multi-accounting | High | Device fingerprint + club graph | Low–Medium |
| Bot play (single account) | High | NSUS behavioural fingerprint + club owner | Medium |
| Bot farm under one operator | High | Device graph + behavioural similarity | Low–Medium |
| Real-time solver assistance | Medium-High | Play-pattern over volume | High |
| External HUDs and overlays | Medium | Client telemetry + ToS | Variable |
Outside-channel collusion sits at the top of the operator's list because, in the long run, a club whose regulars quietly rob the soft players loses those players and stops producing rake, so NSUS and a long-term-greedy owner are aligned against it. But a short-term-greedy owner is not. Chip dumping and embedded bots are even worse, because they can run with the owner's knowledge: the cheat pays rake, the owner collects it, and the only losers are players who will be replaced. The platform and club are not one enforcement team. Their interests overlap on some cheats and diverge sharply on others.
Two enforcers, two sets of incentives
To predict where a behaviour gets caught you have to model two separate parties, not one pipeline. NSUS runs the network layer for the whole population; the owner and his agents run the club layer inside one club. Their evidence, powers and incentives differ.
- Network: behavioural fingerprinting
- Client telemetry on input timing, touch-dwell on mobile, action-confirmation latency, idle behaviour. Cheapest layer, runs continuously, produces a per-session behavioural score. Same signal family GGPoker has refined for years, with a population baseline tuned for the ClubGG demographic.
- Network: statistical play-pattern analysis
- Per-account distributional analysis on VPIP, PFR, 3-bet by position, fold-to-cbet by board texture, bet-sizing histograms, river aggression, all-in equity at showdown. Heavy compute, nightly or weekly cadence. Pure-GTO output is paradoxically easier to flag, because population variability is the baseline.
- Network: account-graph models
- Accounts joined by device fingerprint, IP, deposit method, KYC document, table co-occurrence and action-timing correlations within hands. Multi-accounting and farming fall out here; collusion within a single club is harder because the natural prior is high co-occurrence.
- Club: what the owner can see
- Owners see balance sheets, transfer histories and full table logs for their own club, with tools to restrict seats, expel players, freeze balances and refuse settlement. No algorithm on the network side has this resolution into a single club. The owner can notice a regular winning above their skill envelope days before any statistical verdict exists.
- Club: what the owner will do
- The owner's response is decisive and hard to predict. Catching a cheat costs the owner nothing; acting on one costs him rake, a vouched-for relationship, sometimes a settlement partner. An owner can hold all the evidence and sit on it. From outside the club, a cheat who is known-and-tolerated and a cheat who was never spotted produce the exact same data.
Neither layer subsumes the other. The network layer is broad and slow, good against farms and statistical outliers but blind to context. The club layer sees everything in one club quickly, then acts on whatever the owner's incentives permit. The cheats that thrive are the ones the network cannot prove and the owner does not want to.
The owner as gatekeeper, and as risk
The agent and owner layer separates club apps from public operators. It provides oversight but also creates risk. An agent manages a segment of players: deposits, cashouts, vouching, seat restrictions, transfer visibility. The owner is the agent of agents, the gatekeeper for everyone under him. He is the closest thing ClubGG has to a casino floor.
No model can match his knowledge of this small group: he watches a small set of players for hours, knows who he vouched for, who deposits, who runs hot. A bot that passes NSUS fingerprinting but suddenly wins far above its seat's rate draws his attention quickly, which is the practical argument for gradual volume. He is also the wrong person to trust with that power, because his pay is the rake those same players generate. The honest owner is your best protection on the entire platform; the dishonest one is the hardest cheat to beat, because he is the referee. You cannot tell which you have from the table, only from the chain above it.
This is why running your own club to escape the gatekeeper backfires. The club becomes a visible farm, settlement partners refuse its action, and scrutiny escalates up the union. A modest presence in an established club lasts longer, with normal stack sizes, no table filled with bots, and volume that builds gradually.
Outside-channel collusion
Outside-channel collusion is the dominant problem precisely because it falls between the two enforcers. Colluders share hole cards on WeChat, KakaoTalk or Telegram and coordinate aggression against a target; the platform sees only the visible actions, which look like routine play in isolation, and the owner often has no reason to investigate his own paying regulars. NSUS cannot read the chat, and the owner will not act on the signs. Defence therefore relies on one purely statistical signal, joint behaviour over volume: two players folding dominated hands against each other more than chance, raising in tandem against a third more than position predicts, never bluffing each other on rivers across thousands of hands.
Establishing the baseline is difficult on club apps because of club governance, not mathematics. Standard collusion detection assumes near-random opponent matching, so two players seeing each other for thousands of hands is itself a red flag. Club tables break that assumption because the same regulars sit together every session by design. The correct reference is the per-club joint-behaviour distribution, which means the detector has to learn what normal looks like inside one closed group, and most groups are too small to give it much to learn from.
# Schematic: showdown-correlation detector for club tables
def collusion_pair_score(hands, club_baseline):
pair_obs = defaultdict(lambda: {'soft': 0, 'coord': 0, 'count': 0})
for h in hands:
for p1, p2 in pairs(h.players):
pair_obs[(p1, p2)]['count'] += 1
if held_dominated(h, p1, p2) and not valuebet(h, p1, against=p2):
pair_obs[(p1, p2)]['soft'] += 1
if joint_aggression_vs_third(h, p1, p2):
pair_obs[(p1, p2)]['coord'] += 1
scores = {}
for (p1, p2), o in pair_obs.items():
if o['count'] < 200: # too thin to score
continue
z = z_combined(o['soft']/o['count'], club_baseline['soft'],
o['coord']/o['count'], club_baseline['coord'], o['count'])
scores[(p1, p2)] = z
return scores Production systems condition the baseline on stake level, action sequence, the position relationship between two players and per-pair history length. NSUS can build the per-club priors, yet it lacks the context to interpret a flag inside a club it does not manage; the owner has the context but no incentive to point a collusion detector at his own table. The detector that would actually work needs both, and the two parties who hold the halves do not share a motive. Outside-channel collusion stays unsolved because the parties lack a shared motive, not because of the statistics.
Does the chain go all the way up to GGPoker?
The most consequential governance question is whether a flag inside one ClubGG club travels up to NSUS and across to GGPoker. NSUS does not document its cross-product architecture. An answer must rely on structural inference and the handful of publicly reported cross-product cases. NSUS runs one security organisation over both products, sharing enough infrastructure that any signal collected on either is visible to the same team. The plausible design is a shared abuse-graph ingesting device fingerprints, payment methods, KYC documents and behavioural fingerprints from both. It is reasonable to assume that ClubGG activity is observable on GGPoker through the network-side chain, as with PartyPoker and bwin, or PokerStars and FullTilt. The club-side chain does not work that way. A tolerated cheat known only to one owner never enters NSUS's graph at all, because the one party who could report him is the one profiting from him.
The algorithm versus the owner
The two enforcers demand opposite strategies, and conflating them is the standard error. Against the network layer the right frame is adversarial classification: NSUS builds a model separating bot behaviour from population behaviour, and a well-built bot keeps its timing and bet sizing inside the human range while holding its expected value. The literature runs from Dalvi, Domingos, Mausam, Sanghai and Verma (2004) on adversarial classification and Lowd and Meek (2005) on probing a deployed classifier, through the modern adversarial-ML lineage. This is why pure-GTO output draws attention faster than play with human variety in it: solver-perfect frequencies stand out against a population baseline, while varied timing and sizing read as an ordinary regular. Over a long horizon the account that stays healthy earns more than the one that squeezes the most out of each hand.
The owner needs a different answer. He watches the tables directly and already knows who he vouched for, without running a classifier. Timing variety alone does not settle his questions. Ordinary table conduct does: a stack size like everyone else's, no table filled with bots, and volume that rises over weeks rather than overnight. Choosing the club matters as much as the software, since an owner who runs his own bots or takes a cut from favoured players is a risk to any outsider at his tables. ClubGG's cheating problem depends more on governance than security. The owner remains its strongest detector.
Run the bot on a club table first
Free tests run through Telegram with the Poker Bot AI team, who also send the price. Owner incentives and collusion baselines are fair game in that chat too.