The case against AI solvers in online poker rests on a simple claim: a player running solver software during a hand is no longer competing on skill. The case for solvers points out the same tools have been part of serious training for over a decade, and the better players study them on their own time without dispute. The actual problem sits between those two positions, in the area of real-time assistance, bots that mimic solver output, and the strain those tools have put on the operators that try to police them.
What Solvers Actually Do
A solver is a piece of software that calculates a game theory optimal response to a poker situation. The user inputs stack sizes, bet sizes, board cards, and ranges, and the solver returns a strategy that cannot be exploited by an opposing player who knows the same math. The first widely used solver, PioSolver, came out in 2015 and was followed by GTO+, MonkerSolver, and several others. Each runs on a desktop and produces output that takes minutes to generate per hand.
Used in study mode, solvers tell a player what the balanced action looks like in a particular spot. Used in real time during a live hand, they remove the decision from the player. The first use is study. The second use is cheating.
How Libratus and Pluribus Changed the Conversation
Two academic AI systems shifted how the poker community thought about machine play. Libratus, built at Carnegie Mellon, beat four top heads-up no-limit professionals across 120,000 hands in January 2017 and won the equivalent of $1.7 million in chips. The match took 20 days and ran on a supercomputer at the Pittsburgh Supercomputing Center.
Pluribus, also from Carnegie Mellon and Facebook AI Research, went further. In 2019, it beat five professional players at once in six-handed no-limit Texas Hold’em. The system ran on a single server with 28 cores and used roughly $150 of compute per game. Pluribus achieved a win rate of around five big blinds per 100 hands against the human pros, a margin no human could sustain at that level.
The Pluribus team chose not to release the software publicly. Their published paper described methods in detail, but the working code stayed inside the lab. The reasoning, according to the lead researcher, was that even a marginally improved AI deployed in real money games could drain bankrolls from honest players faster than operators could detect it.
The Real-Time Assistance Problem
Most cheating cases in modern online poker do not involve full bots. They involve real-time assistance, or RTA, where a human player runs solver output on a second screen during live play. The human moves the mouse and clicks buttons. The solver decides what to click. From the operator’s view, the activity looks normal. The mouse moves in human ways, the timing varies, and there is no automation pattern to flag.
Detection has therefore moved from automation signatures to strategy fingerprinting. Operators compare a player’s decisions across thousands of hands against the output of major commercial solvers. If a player’s choices match solver output too closely across enough spots, the account gets flagged for review. Game integrity teams then check session length, mouse patterns, and play frequency to confirm the case.
Joining an Online Poker Game
The choice of online poker game format affects how much of the AI debate touches the average player. Recreational play at low stakes rarely runs into solver-trained opponents, since the cost of using RTA outweighs the small win rate the cheater could extract. Mid-stakes and above is where the policing matters most.
Most operators publish quarterly transparency reports that list how many bot accounts were closed, how much in confiscated balances was returned to honest players, and what detection methods produced the most catches.
© Unsplash
What the Research Community Has Said
The Pluribus paper appeared in *Science* in 2019 and became one of the most cited results in machine poker research. Subsequent work has focused less on beating humans in standard formats and more on extending the methods to multi-table tournament structures, deep stack play, and games with imperfect information beyond poker. A 2020 review of the field pointed out that the techniques behind Pluribus had begun to migrate into auction theory, security games, and negotiation modeling. The poker problem turned out to be a useful proving ground for a wider set of imperfect information research questions.
The reverse migration also matters. As academic work clarified what an unexploitable strategy looks like, beginning with the earlier Libratus result in 2017, commercial solver vendors rewrote their products to produce more practical, faster output. The gap between what an academic system can compute and what a desktop solver can return has narrowed sharply since 2019.
Operator Responses
Major poker rooms have responded to RTA in three main ways. First, they have built detection systems that compare every account’s play to commercial solver output and flag statistical anomalies. Second, they have published rules that prohibit any third-party software running during a session, with permanent bans and balance confiscation as the penalties. Third, several have partnered with the firms that build the most-used solvers to access proprietary models and improve detection accuracy.
The third item draws criticism. Critics argue that handing solver vendors deeper data on the operator’s player base creates conflicts of interest. The vendors counter that detection benefits honest players and that the partnerships are limited to integrity work. The argument has not been settled, and it is likely to continue across the next several rule cycles.
What the Numbers Show
Independent estimates of how much money RTA users move out of the ecosystem run from one percent of total online cash flow at the low end to five percent at the high end. The figure is hard to verify because confirmed cases describe what operators caught, never what they missed. The trend across published transparency reports has been upward in the last three years, a rise that signals more cheating, more thorough detection, or some mix of the two. Most observers in the industry treat the rise as a mix of both.
Players at the lower stakes pay less attention to the issue because the math does not work in the cheater’s favor. At higher stakes, particularly heads-up cash, the share of suspicious accounts has driven some recreational players away from those games. Operators have responded by closing certain heads-up tables and forcing higher-stakes play onto multi-table formats, where solver use is less effective and pattern detection is easier.
The Honest Read
AI solvers have not broken the online poker market. They have changed it. The Pluribus result set the technical baseline for what was possible, and the years since have produced more rigorous study tools, faster strategy improvement among serious players, and harder enforcement of integrity rules. The market also has a small share of accounts that cross from study into real-time use, and the ongoing cost of catching them is now a permanent line item on every operator’s budget.
The question for the next few years is how well detection keeps pace with the next wave of tools. As solvers get faster and harder to fingerprint, and as players who never used RTA before face the temptation when a fast-output mobile solver hits the market, the technical and editorial workload on operators will grow. The market that survives will be the one that pairs strong detection with stakes structures that make cheating uneconomical, and the players who keep showing up will be the ones who trust both pieces are working.
