Poker AI, poker bots, and what happens at the tables

What does "poker AI" mean?

The term gets used for three quite different things, and mixing them up is where most confusion about poker ai starts. First, there are poker bots: automated programs that play hands on your behalf, usually against the rules of any real-money site. Second, there are solvers: offline tools that calculate game-theory-optimal strategy for a given situation, used by serious players to study away from the table, not to play live hands. Third, there are poker AI research projects, most famously Libratus and Pluribus from Carnegie Mellon, built purely to test how far machine decision-making could go against elite human professionals, not to be sold as a product.

When someone searches for "ai poker" or "poker vs ai" in 2026, they're usually curious about one of two things: whether bots are quietly playing against them online, or how AI research stacks up against the best human players. Those deserve separate answers.

GTO Wizard, PokerSnowie and the free alternatives

Two names come up constantly once you start looking into this, and they do quite different jobs. GTO Wizard is a subscription solver platform: it sells pre-computed, game-theory-optimal solutions across common formats and stack depths, so you can look up a spot instead of running a solve yourself, which can otherwise take hours of computing time. PokerSnowie works more like a virtual coach, an AI-driven tool that reviews your hand histories, scores individual decisions against its own solved strategy, and plays practice hands against you directly. Both sit firmly in the study-tool category further down this page, not the live-assistance category that gets accounts closed.

Are there poker bots on real-money sites?

Yes, this happens, though every licensed operator treats it as a serious violation and runs detection systems specifically to catch it. Bots tend to give themselves away through inhumanly consistent timing, betting patterns that never vary with table dynamics, and play that's mathematically sound but strangely flat, missing the exploitative reads a good human makes against specific opponents. Anomaly-detection AI, the same category of tool used for fraud checks elsewhere on a platform, is what most sites lean on to flag suspicious accounts, alongside player reports. If you're playing on a regulated, licensed site, the operator has a direct financial incentive to keep bots out, since a bot-heavy game drives real players away.

Named specifics help here too. GeoComply, a vendor several major poker platforms use, builds detection around device intelligence, checking for the RTA overlays, solver processes and remote-access tools running in the background rather than just watching the cards. WPT Global has been open about running a roughly 50-person game integrity team around the clock, and, tellingly, judges that team on whether flagged cheaters can still sustain a winning rate long-term, not on raw ban counts, since a skilled solver-trained player will look statistically unusual even when they're playing within the rules.

Then there's the grey zone. Real-time assistance software, sometimes called a "poker ai bot" or solver-assist tool, that tells a human player what to do during a live hand sits in a greyer area than a fully automated bot, but nearly every major site bans it under the same real-time-assistance rules. The distinction that matters to most terms of service isn't whether AI was involved at all, it's whether it influenced a decision while the hand was live.

Can ChatGPT play poker?

Not well enough to rely on. The research picture has moved on since Libratus and Pluribus, and the newest results are about general-purpose language models rather than purpose-built bots. A week-long event run by PokerBattle.ai in late 2025 pitted several general-purpose language models against each other in a live no-limit cash game, no code, no solver access, just the model's own reasoning. OpenAI's o3 finished top, ahead of Claude Sonnet 4.5, with Grok and Meta's Llama trailing well behind, and the spread between them was wide enough to show these models still play with very different levels of risk discipline rather than converging on one style. Separate academic work on a system called ToolPoker found something similar: language models can explain solid poker reasoning in words, then take the opposite action anyway, a mismatch researchers call the "knowing-doing gap." Giving the model access to a real solver mid-decision closed most of that gap, but that setup is exactly the real-time-assistance category described above, which is why the researchers behind it argued against ever commercialising it for live play.

When did poker AI beat the pros?

Two results are worth knowing if you want the real history rather than the marketing version. In 2017, Libratus beat four top heads-up no-limit specialists over 120,000 hands, a result significant enough that it settled a long argument about whether no-limit poker's imperfect information made it too hard for AI to solve cleanly. In 2019, Pluribus went further, beating a table of professionals in six-player no-limit hold'em, a much harder multiplayer setting where classic game theory guarantees mostly break down. Neither project became a consumer app. They were research demonstrations, and the techniques behind them fed into the solver software serious players now use to study, not into anything you'd play against directly online.

Is using AI to study poker allowed?

Yes, between sessions. The accepted rule, across most regulated sites, is simple: study with AI between sessions, never during a hand. Running a solver, a hand-history reviewer, or the Chen formula checker and equity calculator on this site before or after you play is normal and expected at a serious level. Having any tool open, generating advice while a hand is live, is what gets accounts closed. Heads-up displays showing historical stats are a murkier case, so it's worth reading a specific site's terms rather than assuming, but the study-versus-live distinction covers the large majority of cases. For a broader look at where AI shows up around poker and other games, the glossary is a good next stop.

Worked example: K-Q suited against K-9 suited

Take King-Queen suited and King-Nine suited, two hands a newer player might rate about the same, since both start with a King. The Chen formula treats them very differently. King-Queen suited: the King scores 8 points, add 2 for being suited, add nothing for the gap since the two cards are directly connected, for a final score of 10, comfortably a raise-from-anywhere hand. King-Nine suited: the King still scores 8, add 2 for suited, but the three-rank gap between King and Nine costs 4 points, dropping the total to 6, a hand most charts treat as speculative and position-dependent rather than an automatic raise. Same top card, same suited bonus, a four-point swing purely from how connected the two cards are, exactly the kind of gap a quick mental estimate misses and the calculator catches instantly.

Where to start

Poker is the one casino game where skill decides who wins over time, so practise in the right order.

• Start with preflop discipline: score your hands with the Chen formula calculator.
• Check close calls against the price with the equity calculator, and use ICM for tournament spots.
• Study with solvers away from the table only. Real-time assistance breaks the rules on almost every real-money site.