My take on poker AI at BetOnline.

What we mean by "poker bot" in 2026

The word "bot" is overloaded. For the purposes of these notes, a poker bot is software that reads visible game state — your own hole cards, the community cards, the position and stack of each opponent, the betting sequence — and produces an action under a real-time latency constraint. It does not read packets. It does not know what opponents hold. It does not know the order of the deck. Everything a poker bot does, a focused human could do with the same inputs, given enough time and consistency. The bot's advantage is doing it for ten hours, on six tables, without fatigue, with sizing and frequency lifted from a solver.

Underneath that definition sit three layers. A solver-anchored baseline — pre-computed strategies from counterfactual regret minimisation (CFR) variants, usually PioSolver or GTO+ for heads-up and 6-max trees, MonkerSolver for multiway. These outputs are far too large to query naively and get compressed into state-and-action abstractions keyed on bucketed board textures. An opponent model, where most of the current research happens: online updates per hand against a known opponent, with archetype priors for new ones. BetOnline keeps stable screen names and permits HUDs, so the opponent-model picture here is meaningfully different from GGPoker — the long-horizon HUD attack still works. And a policy combiner that chooses how far to deviate from the baseline given the current opponent estimate, with behavioural noise on top to keep timing and sizing distributions inside the population envelope.

BetOnline and the Chico Network

BetOnline began in the early 2000s as a sportsbook and casino brand, with poker added later as a third vertical. The operator runs out of Panama under a Panama gaming licence (with a separate Curaçao licence covering parts of the casino and sportsbook product), and the poker client sits on what is now called the Chico Network — a small private network whose poker stack is derived from older PartyPoker software. The Chico Network also includes sibling sportsbook-branded rooms; players see one liquidity pool across them.

Three structural facts matter for anyone studying this room as an automation target. First, the codebase lineage: a PartyPoker-derived client patched over many years rather than rebuilt. That codebase has known historical edges (UI quirks, predictable lobby behaviour, older anti-fingerprinting on the client) and lacks some of the modern hardening you would see in a GGPoker client. Second, the player pool: post-Black-Friday US and Canadian recreational players, supplemented by South American liquidity. PokerStars and Full Tilt exited the US market on 15 April 2011, and BetOnline, ACR and Ignition absorbed the leftover demand. The pool has stayed weaker than European regulated rooms for over a decade. Third, the HUD policy: third-party HUDs and tracking software (Holdem Manager, PokerTracker) are tolerated in practice on BetOnline. Player IDs are stable. The long-horizon data-mined HUD that died at GGPoker still functions here.

How detection works at a smaller operator

The detection topology at BetOnline has the same logical shape as at any large poker operator — behavioural fingerprinting at the client, statistical play-pattern analysis offline, an account graph for multi-accounting and collusion, and a human review layer that signs off the actual bans. What differs is the budget. The security team is smaller. Offline batch jobs run on a less aggressive cadence. Review-queue capacity is lower. The practical consequence is that the system is more reactive than proactive: the rule-of-thumb pattern is that a quietly-running single-account bot at small or mid stakes survives until a triggering event — a large withdrawal, a player complaint, an anomalous winrate over a long sample — pushes the account into the human-review queue.

Behavioural fingerprinting

Client-side telemetry on input-timing distributions, mouse-path geometry on desktop, touch dwell on mobile, idle behaviour between hands. The Chico client collects this at a lower fidelity than newer operators. Naive constant-latency bots still get flagged; carefully shaped behavioural noise gets through reliably.

Statistical play-pattern analysis

VPIP/PFR distributions, bet-sizing histograms, fold-to-3-bet by position, all-in equity at showdown. The signal is the same as at GGPoker but the offline cadence is slower; outlier accounts can sit in the long tail of a distribution for months before a human looks at them.

Anti-collusion graph models

Aggressive on the multi-accounting side. Account graphs joined by IP, device fingerprint, deposit method (especially shared crypto wallets), and table co-occurrence catch the dumping rings that hurt the room financially. This is the layer where the operator has incentive to spend, and they spend.

Human review on flagged accounts

The decisive layer everywhere. Volume is the differentiator at BetOnline: fewer reviewers, longer queue, slower cycle. The 2014 and 2018 bot cleanups were both batched human-review actions triggered by external complaints and forum pressure, not realtime detection firing.

What this site covers

Open questions for developers

The notes on this site are work in progress. Five threads that I think are genuinely open against the BetOnline target specifically:

1. Long-horizon HUD as opponent prior. Stable screen names and tolerated HUDs mean a data-mined long-horizon profile is available on most opponents. How much that prior actually helps an online model that already converges in 100-200 hands is empirically unclear once you control for the soft-pool baseline.
2. Codebase-specific UI automation. A PartyPoker-derived client has years of accumulated UI quirks: lobby state machine edge cases, multi-tabling layout assumptions, table-redraw behaviour after disconnects. None of this is in any solver paper. The boring engineering matters more here than the math.
3. Reactive detection budgeting. If the operator's enforcement is bursty (quiet for months, then a public-pressure-driven sweep), what is the right account-lifetime risk model? The standard "budgeted detection probability per account" frame assumes a stationary classifier, which is not the empirical pattern here.
4. Crypto-bankroll segmentation. Bitcoin, Ethereum and stablecoin rails make deposits and withdrawals cheap and fast, but they also leave on-chain trails that operator anti-collusion systems use to join accounts. The right wallet hygiene is non-trivial and is a vector by which bot farms get rolled up.
5. MTT and SNG ICM at the Sunday-flagship scale. BetOnline's Sunday tournaments and Jackpot Sit & Gos have ICM dynamics that the cash engines do not encode. Whether to ship an MTT-aware engine for an operator with a smaller MTT volume than PokerStars is a product question, not a pure research question.

Corrections and contributions welcome. The chat link below is read by the team; serious questions get answered.

Where these notes are useful

If you are evaluating BetOnline specifically — its Chico Network codebase, its reactive detection posture, its tolerated-HUD policy, and the soft post-Black-Friday US pool — the practical value here is calibration: separating what the room's older PartyPoker-derived stack actually changes from the forum folklore around it, and grounding account-lifetime expectations in the documented 2014 and 2018 enforcement pattern rather than in marketing claims. That calibration is the point of the site. If you are working on something concrete against this target and want a sober second read on detection risk, data, or architecture, reach the team — the channel is low-volume and read by people who build this software, not a sales desk.