AI Poker Bots: What They Can Beat and How Sites Catch Them

In October 2026, players on the Two Plus Two forum flagged 2,587 suspected bot accounts at QQPK, and the site banned 724 accounts for automated play in its first sweep. Two days later, GGPoker banned all third party poker software and added a check that flags players whose decisions are right far too often against cards they should not be able to see.

Poker AI is not new. A University of Alberta program called Polaris beat a team of limit hold'em pros in Las Vegas in 2008, and Carnegie Mellon's Libratus won $1,766,250 in chips from four heads up no limit pros in 2017.

This guide to AI poker bots separates what AI can do in a research match from what actually happens at online tables, with every case, ban and refund backed by a named source. VIP-Grinders has covered online poker since 2013, and this guide is written by players who have spent years at the tables these bots target.

This page covers:

A poker bot is software that plays online poker by itself. It reads each hand, chooses an action and clicks the buttons, with no person making the decisions.

Online rooms ban them. WPT Global's rules prohibit “robots or any other external player assistance (EPAs) program”, and GGPoker and ACR say every decision at the table must be the player's own.

How a Poker Bot Plays

Every bot does the same three jobs, hand after hand, at every table it sits at:

Rooms go after each of those steps. According to Card Player, ACR runs regular client updates to disrupt bots that read pixels, and PokerStars said in 2023 that monitoring mouse clicks had helped it catch fully automated bots.

Bot makers know this. Matt Mazur, a developer who ran his own bot on PokerStars for nearly two years, wrote in 2009 that he modelled its clicks on his own clicking pattern so it would not hit the same spot every time.

Poker Bots vs RTA

Real-time assistance, or RTA, is anything that helps a player make decisions while a cash game or tournament is in progress. PokerNews lists solvers, range calculators, push or fold charts and preflop charts based on stack sizes or earlier action among the tools that count.

The difference is who plays the hand. With RTA, a real person still sits at the table and clicks every button while following outside advice, but a bot plays alone.

Both break the same rule. GGPoker's terms say “Any decision made at the poker table must be made solely by you”, and ACR's statement on an RTA case in September 2026 said “Every decision at an ACR Poker table must be made by the player.”

Studying with a solver away from the tables is allowed, and PokerNews notes you can use one as much as you like when you are not playing. Check each room's rules first, because GGPoker now requires all third party poker software and automation tools to be closed before you open its client.

Some well known players have been caught using RTA, and our page on the biggest RTA cheating cases covers what happened in each one.

Yes. AI programs have beaten elite professionals at heads up limit hold'em, heads up no limit hold'em and six player no limit hold'em, all in research matches run by universities and Facebook AI.

The games got harder as the programs improved, from limit hold'em against one opponent to no limit against five. The table lists every landmark match, and the sections below explain what each one proved:

Polaris and the First Win at Limit Hold'em

The University of Alberta's Polaris took on two pros at the first Man-Machine Poker Championship in July 2007, held at an artificial intelligence conference in Vancouver. Phil Laak and Ali Eslami played duplicate limit hold'em against it, where each human was dealt the cards Polaris held in the other room. The humans won two of the four sessions, lost one and drew one.

A year later, Polaris 2.0 played a team of pros from Stoxpoker at the Rio in Las Vegas, during the World Series of Poker. It won two of the four matches there, lost one and drew one, and Guinness World Records lists it as the “First poker bot to beat human professionals in a live tournament”.

Has Texas Hold'em Been Solved?

Only one version, and only essentially. In January 2015, University of Alberta researchers published a paper in Science titled “Heads-up Limit Hold'em Poker is Solved”, built on a program called Cepheus.

Essentially solved means that even a lifetime of play could not show its strategy is flawed, with 95% confidence. In plain terms, no human could play enough hands against it to prove it makes a mistake.

No limit hold'em has not been solved. Carnegie Mellon puts the number of situations a player can face in heads up no limit at 10¹⁶⁰, a 1 followed by 160 zeros, and even Libratus could not calculate a perfect strategy, so it worked with an approximation.

Libratus and DeepStack Take On No Limit Pros

No limit is a much bigger game, because players can bet any amount. Carnegie Mellon's first attempt, Claudico, played 80,000 hands against four pros at Rivers Casino in Pittsburgh in 2015, and the humans finished $732,713 ahead in chips. CMU said that margin was too small to be statistically significant, so the match counts as a draw.

Two years later, Libratus, built by Tuomas Sandholm and Noam Brown, went back to the same casino:

Around the same time, DeepStack, built by the University of Alberta with Charles University and the Czech Technical University in Prague, took on 33 pros from 17 countries in late 2016. Eleven of them completed 3,000 hands each, and DeepStack beat 10 of those 11 by a statistically significant margin, winning 49 big blinds per 100 hands overall.

Pluribus Takes On Five Pros at Once

Every program above played one opponent at a time. Pluribus, built by Carnegie Mellon and Facebook AI and published in Science in July 2019, beat elite pros at six player no limit hold'em:

Pluribus's creators did not release the code: Noam Brown told MIT Technology Review it “could be very dangerous for the poker community”.

A solver is a program that calculates a balanced strategy for a poker situation from the inputs you give it, such as the ranges, stack sizes and bet sizes. Unlike the research programs above, solvers do not play: they show a player what to do, and since 2013 they have changed how professionals study the game.

The Private Solvers Sold to High Stakes Pros

Before 2013, the poker software on sale was limited. GipsyTeam reports that Cardrunners EV cost $50 and could not meet the needs of high stakes players.

By January 2013, Oleg Ostroumov, who had studied engineering and maths in Moscow, had built a no limit hold'em solver that worked out heads up play from the flop with set bet sizes. By his own account, he sold exclusive access to a group of six players led by Alex “Kanu7” Millar for $200,000, paid over about 18 months.

He later sold an Omaha solver for $300,000, according to GipsyTeam. Ostroumov describes his first program as the first no limit hold'em solver used by high stakes professionals.

Commercial Solvers Reach Every Player

Full solvers went on public sale two years later. PioSOLVER, created by Polish programmer and player Piotr Lopusiewicz, went on sale in 2015 for $249, and MonkerSolver followed in February 2017 with Omaha and pots of three or more players:

Browser tools such as GTO Wizard came next, with libraries of presolved spots that work on any device. Bernard Marr writes that competition at higher stakes has become much stronger since solvers like PioSOLVER arrived.

Our guide on how to study with a solver shows how to use one, and our tools hub covers the solvers and trackers we have tested.

From Study Tool to Cheating Aid

Once solved spots could be looked up in seconds, the risk moved from study to the table. A player could check a spot while the hand was still being played, which is exactly what rooms ban as RTA.

GTO Wizard now helps rooms catch it:

Lookup checks are only one part of how rooms catch cheats, and the detection section further down covers what each one says it uses.

The newest AI players are chatbots: general purpose language models such as ChatGPT, Claude, Gemini and Grok, which were not built for poker. Since 2025, a handful of public tests have put them at the table.

These are the public tests we found, with the results each source reported:

PokerBattle.ai and the Chatbot Poker Contest

PokerBattle.ai, built by former product manager Max Pavlov, sat nine chatbots at three $10/$20 no limit hold'em cash tables for five days in October 2025, using play money. Each started with a $100,000 bankroll, and OpenAI's o3 finished on top, $36,691 up over 3,799 hands, ahead of Claude Sonnet 4.5 and Grok 4.

Meta's LLAMA 4 lost its whole bankroll after 3,501 hands, voluntarily putting money in the pot preflop more than 60% of the time, which PokerNews put at nearly three times the rate of the others. Pavlov himself warned against reading too much into the results, given the small sample.

Kaggle Game Arena and the Google DeepMind Exhibition

In February 2026, Google DeepMind and Kaggle added heads up no limit hold'em to Game Arena, their benchmark where AI models compete, with Nick Schulman, Doug Polk and Liv Boeree commentating. The top eight models on the leaderboard played over three days, and the final on February 4 was all OpenAI: o3 against GPT-5.2. GipsyTeam reports that GPT-5.2 won it, at 35 big blinds per 100 hands.

Phil Galfond vs Grok

In late October 2025, Phil Galfond and xAI's chatbot Grok agreed in principle, in public on X, to a heads up match. The reported terms were:

Galfond said “As long as it's a fair platform, I'm in, and happy to stream”, but no platform was agreed. Days later, Grok told Todd Witteles, who was trying to set up his own match with it, that it had overstepped and that xAI does not do poker side bets.

No match has been reported since, and our roundup of the Galfond vs Grok prop bet and other famous wagers covers the full story.

Chatbots That Write Their Own Poker Bots

Husky Hold'em Bench, presented at a NeurIPS 2025 workshop, tests something different. Each chatbot writes its own poker bot, the bots then play each other in six player tournaments, and the bot written by Claude 4 Sonnet topped the leaderboard.

It matters for online poker because it measures whether a chatbot can write the software that plays, not just play a hand itself. All three completed tests above pitted AI against AI, with no human players at the table.

Online poker bots are often run in groups, known as bot farms or bot rings, with one operator controlling many accounts at once. Working together lets a ring do things a lone bot cannot, such as playing several seats at the same table.

How a Bot Farm Operates

The published cases show four main kinds of bot operation:

Ignition responded to the January video by saying the footage was from 2022 and that every account shown, along with others found in a review, had been deleted.

The Poker Bot Enforcement Record

Rooms do catch bot rings, and some publish what they found and what they paid back. These are the published cases we could confirm with named sources:

Players often raise the alarm first: Two Plus Two users flagged the QQPK accounts, a post on X put the Ignition video in front of the community, and players tracking the PokerStars accounts produced the 2015 estimate. Podcasters keep pushing the issue too, and our profile covers Joe Ingram's bot investigations.

Are There Bots on Online Poker Sites Today?

Yes, and the people who fight them say so openly. John Andress, head of game integrity at A5 Labs, told PokerNews in 2024 that operators who claim nobody uses bots or RTA on their platforms “do not understand the evolution of the problem, or they are being dishonest”.

Bloomberg's Kit Chellel, who investigated Russian bot operations, put it more bluntly in a 2024 interview with PokerNews:

“It is simply impossible to have a poker site without bots.”
Kit Chellel, Bloomberg

He added that rooms can limit them but cannot prevent them, and the QQPK case in October 2026 shows how large a ring can still grow before it is caught. Rooms differ in how quickly they catch bots and what they do with the money, and the next section covers what each room says it uses.

Do Poker Bots Actually Make Money?

Some do, until they are caught. Players estimated that the suspected PokerStars ring won almost $1.5 million before accounts were closed, and Chellel noted that detection takes many hands, while burner accounts keep coming on, using bots and leaving again.

What happens to the money after a ring is caught depends on the room:

For anyone tempted to buy a bot, the account and the balance in it are what is at stake: in most of the cases above, rooms closed the accounts and paid money back to the players who lost to them.

Rooms keep most of their bot detection secret, and CoinPoker says it holds back the details so it does not hand tips to cheats. What they have said in public still shows how online poker sites detect bots: software flags suspicious accounts, and a security team decides what happens next.

The Signals Security Teams Watch

The measures the rooms have described point to five kinds of signal:

GGPoker's check on decisions is the newest of these, added in October 2026 alongside its existing checks for collusion, ghosting, botting and RTA. Our article on GGPoker's ban on third party software covers everything that changed.

What Each Room Says It Uses

These are the measures each room has described in public. The detection figures are the rooms' own, and we found no independent audit of them:

Our articles on how WPT Global detects and bans bots and ACR's layered bot defences go into more detail on those two rooms.

What Happens When a Bot Is Caught

The rooms describe the same broad steps:

Players help too. WPT Global's Rewarding Integrity in Poker scheme had paid more than six figures to several hundred players for reports by February 2026, PokerNews reported, and the protection section below covers how to make a report a room can act on.

Not all AI at the poker table is cheating. Some rooms and broadcasters now use it in the open, with rules that keep it fair for everyone:

ACR Skip-It Tournaments

Skip-It tournaments went into beta at ACR in March 2026, costing $1.05, $5.25 or $10.50 to enter. Each one runs in three stages:

Players can also buy straight into Day 2 for $26.25, starting with fewer chips. Our article covers how Skip-It tournaments work in more detail.

The Bluff Reading Model on the WSOP Broadcast

ESPN's 2026 World Series of Poker coverage, produced by Peyton Manning's Omaha Productions, showed an AI model that estimates whether a player is bluffing or holding a strong hand. It was built over about six months by independent AI engineer Luke Geel, and it reads body language such as posture, facial movement and blink rate rather than the cards.

The reads are shown only after the player is out of the hand, so they cannot affect live play. Geel also said he saw no reason yet to run it on Daniel Negreanu, who has probably worked hard to remove his tells.

Players were divided. Quoted by PokerListings, Sam Greenwood called the model “likely over or under fit and not very good”, while Ian Simpson warned of the “scope for abuse off of the feature table” if it works.

Our article on WSOP AI bluff detection covers the tool and the debate around it.

You cannot stop a bot from sitting at your table, but you can make sure the room hears about it and choose rooms that act. Most of the work happens on the room's side, so your part is to notice, report and pick your sites with care.

Can You Tell if a Player Is a Bot?

Sometimes, but it is harder than many guides suggest. SpadePoker lists three signs to watch for:

Those signs are getting weaker. Andrew Kuznetsov, a poker AI developer whose team builds bot detection, argued in August 2026 that modern bots vary their timing with how hard the spot is, stop sessions at set times or win rates and copy the bet sizes regulars use, while heavy multi tabling is now normal for human pros.

The bigger gap is what each side can see. You see one table, while the room's server sees every hand an account plays on every table, along with its devices, IP addresses and links to other accounts.

None of these signs is proof on its own. Strong human regulars can look robotic too, so leave the verdict to the room and send it the evidence.

How to Report a Suspected Bot

A report the room can act on is short and factual:

Some rooms reward good reports. WPT Global accepts anonymous tips, but you need to give contact details to be eligible for a reward, and affected players are compensated first when a case is confirmed.

Choose Rooms That Publish Their Enforcement

Bloomberg's Kit Chellel said players cannot tell which sites keep bots out best, or whether the opponent in front of them is a bot: “You can't know that.” What you can check is whether a room tells you what it catches:

Publishing these numbers does not make a room bot free, but it shows the room is catching bots and paying players back.

The fear that AI will kill online poker is older than ChatGPT. Since Pluribus beat the pros in 2019, the warnings have come from the team that built it, from an investment bank and from a Silicon Valley investor:

Why the Warnings Were Taken Seriously

Pluribus did not need the supercomputers that earlier game playing AI relied on, and Young called that a major development because it made the strategy affordable and quick to copy. Morgan Stanley cut its price targets for three gambling stocks, with The Stars Group, then the owner of PokerStars, the most exposed to online poker.

Palihapitiya went further. He argued that AI agents would make money for their creators on legal sites, losing some hands on purpose so they pass as normal players.

What Has Happened Since

The industry's answer at the time was that bots were nothing new. Severin Rasset, then managing director of poker at The Stars Group, told iGaming Business in 2019: “Bots and attempts at manipulation have both existed since day one of online poker.”

So far, the record looks like this:

SharkScope counts seats rather than people, and our monthly online poker traffic report tracks those numbers network by network. The threat has not gone away, though: as the protection section shows, bots are already getting harder to spot by eye.

What AI Changes for Players Who Want to Win

Palihapitiya gave online poker a few years, and two years on the tables are still busy. The bigger change is in what it takes to win at them, as our Head of Poker Content puts it:

“AI won't kill online poker, but it has already changed what it takes to win. Every year the edges get thinner, and they will keep shrinking as more players get access to better tools. Rooms can block bots, but they can't stop players using AI to study, so the players who keep winning will be the ones who find an extra edge somewhere else.”
Mark Patrickson, Head of Poker Content

Whether that edge is still there to find comes down to skill, volume and rakeback, and our guide on whether online poker is still profitable runs the numbers.

Rakeback is the one part of that edge you can secure before you play a hand, because the deal is set when you sign up. Our poker rakeback deals show what each room pays back, including exclusive offers and private freerolls for VIP-Grinders players.