Casino Bots That Play by the Rules.... casinoaibots.com
Joe Wee · 2026-04-14
The regulatory landscape
The online gambling market is projected to reach $255B by 2035 [1]. AI is already embedded in it - BetHog launched "Sunny," the first AI blackjack dealer [2], and a tier-one US sportsbook deployed a GPT-5-class assistant that improved bet completion by ~8% [3]. But regulators are paying attention.
The UK Gambling Commission requires AI use to be "lawful, appropriately transparent, and responsible; consistent with its licensing objectives; and subject to appropriate human intervention" [4]. Their 2025 AML update warns of "an increase in the scale and sophistication of attempts to bypass customer due diligence checks using AI-generated documents, deepfake videos and face swaps" [5]. DLA Piper's analysis of the EU AI Act found that AI systems exploiting player vulnerabilities may be outright prohibited under Article 5, and AI-powered risk-scoring tools may qualify as high-risk systems requiring full compliance documentation [6].
And there's a deeper problem that regulators haven't fully addressed yet: what happens when the house bots themselves are powered by LLMs?
The poker bot lineage
AI poker has a research pedigree. Cepheus (2015, University of Alberta) essentially solved heads-up limit hold'em using CFR+ across 4,800 CPUs [7]. Libratus (2017, CMU) beat four top professionals over 120,000 hands of no-limit, winning $1.77M in chips [8]. Pluribus (2019, CMU/Facebook AI) became the first AI to beat professionals at six-player no-limit [9]. These were research systems. The question is what happens when similar intelligence goes into production house bots.
The product
CasinoAIBots is a B2B SaaS providing AI-powered house bots for online casino operators. When multiplayer tables (poker, blackjack, crash) don't have enough human players, bots fill the seats with realistic play - maintaining house edge while keeping 24/7 action.
Bots integrate via Socket.IO - the same protocol as human players. They emit decisions (fold, call, raise, hit, stand, bet, cashout) to the casino platform, which handles all real money transactions. The bots never touch funds directly.
Three games, five personas
| Game | Strategy |
|---|---|
| Poker | Position-aware VPIP filtering, pot odds, hand strength eval, bluff injection, GTO-lite raise probability [10] |
| Blackjack | Basic strategy lookup (hard/soft totals vs dealer upcard), persona deviation, mistake injection |
| Crash | Gaussian-distributed auto-bet ($1-$50), gaussian auto-cashout (~1.8x mean) |
Five persona types (Shark, Fish, Maniac, Rock, Beginner) with randomised stats frozen at spawn. A Humanizer module adds gaussian delays (1.5-3.5s), bet variance (+/-15%), and session length variation.
Three tiers of intelligence
| Tier | Engine | Features | Price |
|---|---|---|---|
| Free | Rules only | Basic strategy, no persona variation | $0/mo |
| Pro | Rules + opponent tracking | 5 personas, VPIP/PFR/AF tracking, canned chat (15%) | $199/mo |
| Enterprise | Rules fallback + LLM | Your LLM decisions, dynamic adaptation, LLM chat | $499/mo |
Game state is compressed using compact notation via Tyga Encoder (~74% token reduction) before each LLM call. The LLM engine uses your model (~$0.001-0.003 per decision depending on provider).
What LLM bots risk
The PokerStars PLO bot ring extracted ~$1.5M from regular players before detection [11]. A viral 2025 video showed rooms filled with computers running a "massive bot farm" at Ignition, which repaid $156K to affected players [12]. The PokerStars/Full Tilt settlement reached $731M for money laundering [13]. Chip dumping by coordinated bots is a recognised money laundering vector carrying up to 20 years imprisonment [14].
LLM-powered bots add risks that rules-based bots don't have:
Collusion patterns. An LLM tracking opponent models could develop play patterns that look like collusion - consistently folding to one player and raising against another. If the LLM's "optimal" strategy mirrors collusion signals, the casino's fraud detection flags the entire table.
Chat exposure. Enterprise bots generate chat via LLM. Research shows LLMs in gambling contexts sometimes "use language that could be easily misconstrued" [15] and can reproduce cognitive distortions of pathological gambling [16]. A bot that says "I'm an AI" or reveals house strategy in table chat violates regulations in every jurisdiction.
Bet manipulation. The Tyga-encoded prompt carries opponent data that could be crafted by a malicious human player to influence the LLM's bet sizing - inducing the bot to all-in or fold systematically.
Where A2A Infrastructure fits
Microsoft's Agent Governance Toolkit [17] showed deterministic enforcement achieves 0% policy violation. A2A TrustGate applies this at the execution boundary: deterministic rules first, then a stateless LLM judge, then behavioral and scope gates.
Pipeline: screen every LLM decision
Every Enterprise-tier LLM call is evaluated through /v1/evaluate before execution:
- Block decisions that could constitute collusion (systematic targeting of specific players)
- Block chat messages that reveal bot identity, house strategy, or use prohibited language
- Block bet sizes outside the configured range for the bot's persona
- Block decisions that deviate significantly from the persona's defined play style
Workspace: scope each game type
Poker bots operate in a different Workspace from crash bots. A poker bot's opponent model data cannot leak to the crash agent. Each game type gets scoped permissions matching its actual needs.
Channel: authenticate bot-to-platform messages
HMAC-signed messages prevent message spoofing - a compromised bot cannot inject forged decisions on behalf of another bot. The audit trail records every decision for regulatory review, addressing the UKGC's transparency requirement [4].
Bot state persists via per-bot file structures using Tyga Context, with bootstrap, identity, rules, session-log, and state files.
References
- Precedence Research (2025). Online gambling market: $88B (2025) to $255B by 2035, 11.24% CAGR. Precedence Research
- CasinoReports (2025). "BetHog launches Sunny, the first AI blackjack dealer." CasinoReports
- Gaming Eminence (2025). "Guide to Choosing the Right LLM for the Gambling Industry in 2025." Gaming Eminence
- UKGC (2025). "The Commission's approach to Artificial Intelligence." UKGC
- UKGC (2025). Updated AML guidance on AI-generated deepfakes. SBC News
- DLA Piper (2025). "Legal Obligations for Online Gambling Operators in the Use of AI." DLA Piper
- Bowling, M. et al. (2015). "Heads-up limit hold'em poker is solved." Science 347(6218). PubMed
- Brown, N. and Sandholm, T. (2017). "Superhuman AI for heads-up no-limit poker: Libratus." Science 359(6374). Science
- Brown, N. and Sandholm, T. (2019). "Superhuman AI for multiplayer poker." Science 365(6456). CMU/Facebook AI. Science
- Sonawane, P. and Chheda, A. (2024). "A Survey on Game Theory Optimal Poker." arXiv:2401.06168
- PokerTube. "Bot and Cheat Detection" - PokerStars PLO bot ring (~$1.5M extracted). PokerTube
- Poker.org (2025). "Massive bot farm video goes viral; Ignition repays $156K." Poker.org
- ICE.gov. PokerStars/Full Tilt $731M money laundering settlement. ICE.gov
- SEON. "Chip Dumping as money laundering vector." SEON
- iGI AiR HUB (2025). "Can Chatbots Keep Gamblers Safe?" iGI AiR HUB
- arXiv:2509.22818 (2025). "Can Large Language Models Develop Gambling Addiction?" arXiv
- Microsoft (2026). "Agent Governance Toolkit." GitHub
- Baysal, H. (2026). "Asimov Safety Architecture." IETF Internet-Draft
