In 2019, a piece of software sat down at a six-player no-limit Texas hold’em table and did something no machine had done before. Pluribus, a collaboration between Facebook’s AI lab and Carnegie Mellon University, beat elite human professionals across thousands of hands. The professionals were not amateurs or mid-stakes grinders. They were among the best players alive. The bot won at a rate of 48 mbb/game, a margin considered very high by any professional standard. Translated into dollars, Pluribus would have earned roughly $1,000 per hour if each chip carried a dollar value.
The cost to train this opponent was about $144.
That number should stay with you. A hundred and forty-four dollars produced a system capable of outplaying humans who had spent decades refining their craft. The gap between production cost and destructive capability defines the problem poker platforms still face today.
What Pluribus Actually Did
Pluribus played six-max no-limit hold’em, a format with incomplete information and multiple opponents. Earlier poker bots had solved heads-up games, where two players face each other. Adding more players increases complexity in ways that scale badly. Each additional opponent multiplies the possible game states. Pluribus handled this by computing strategy in real time rather than relying on pre-calculated responses for every situation.
The bot searched through possible actions and outcomes during play, adjusting to what opponents were doing. It bluffed. It varied bet sizes. It mixed strategies to remain unpredictable. When researchers examined its play, they found moves that surprised them, decisions no human had considered optimal before.
The Cost of Playing Against Code
When players choose to play poker online, they enter rooms where detection systems run constant scans for automated opponents. PartyPoker shut down 291 fraudulent accounts in 2024 and returned $71,771 to affected users, a rise from 214 closures and $32,433 the year before. 888poker paid out over $250,000 in 2024 to compensate victims of bots and real-time assistance tools. The numbers show enforcement working, but they also confirm the threat remains active at every stake level.
The Underground Followed Quickly
Pluribus was never released to the public. The researchers understood what would happen if the code circulated. That caution did nothing to stop the proliferation of commercial and custom-built bots in the years that followed. The methods Pluribus demonstrated became a blueprint. Developers with enough skill could approximate its approach, and many did.
In January 2024, a botfarm operating on the WPN Network was exposed after players on the 2+2 poker forum investigated suspicious patterns. The operation had extracted $10 million before being caught. Separate reporting uncovered a large-scale bot operation in Siberia that operated openly under the name Bot Farm Corporation.
These are the cases that became public. Detection relies on behavioral analysis and pattern recognition. Bots designed to mimic human timing, mouse movements, and decision variance can avoid flags for months or even years.
Why Detection Remains Difficult
A human player at an online table clicks buttons, moves a cursor, and takes time to think. Software can replicate all of these behaviors. Modern bots introduce randomized delays between actions. They vary bet sizing within reasonable ranges. They fold strong hands occasionally to avoid statistical anomalies. The better ones play a strategy that looks like a solid regular, not a perfect machine.
Platforms use data mining to catch outliers. Accounts that play 20 tables simultaneously with identical response times attract attention. Accounts that run a win rate far above human norms raise flags. But a bot running four tables with a modest edge at mid-stakes can blend in for extended periods. The economics favor the cheater: low cost, passive income, and replaceable accounts if one gets caught.
Since 2018, PartyPoker alone has closed over 2,540 fraudulent accounts and returned more than $2 million to affected players. The scale of enforcement suggests the scale of the problem.
How Players Respond
Recreational players often remain unaware of the threat. Professionals and semi-professionals know the reality and adjust. Some move to live poker, where bots cannot operate. Others focus on formats or stakes where bot activity appears lower. A portion has quit entirely, calculating that the expected value of online play has degraded below acceptable thresholds.
Forums and private communities share suspicions about specific accounts. Players track hand histories and compare them against known bot profiles. This grassroots enforcement supplements platform efforts, though it lacks the data access necessary for definitive identification.
The Economics Have Not Changed
The base Pluribus strategy cost $144 to produce. Modern computing has become cheaper since 2019. Cloud services offer processing power at rates that make training sophisticated models accessible to small operations. A bot that earns $500 per month at low-stakes tables pays for itself quickly and can run indefinitely with minimal oversight.
Platforms face a calculation of their own. Aggressive enforcement costs money and risks alienating some portion of the player base, including legitimate grinders who get flagged by overly sensitive systems. Insufficient enforcement lets bots drain the ecosystem, driving away recreational players who provide the money everyone else competes for.
Where This Leaves the Tables
The Pluribus paper demonstrated that six-player no-limit hold’em had been solved well enough for a machine to beat humans. The implications spread faster than any safeguard could contain them. Every online table now operates under the possibility that one seat belongs to code rather than a person.
The arms race continues. Detection improves; evasion improves. Money changes hands in both directions, returned to some victims while others never know they lost to a machine. The game itself remains the same 52 cards, the same betting rules, the same mathematical edges to find. The opponents have changed. Some of them never blink.
Conclusion
Pluribus was not an isolated experiment. It demonstrated that the technical barrier to outperforming humans in complex, multiplayer poker environments had already fallen. As computing costs dropped and strategies became easier to replicate, the incentives for abuse only grew stronger.
Online poker still functions, but it does so under constant pressure. Every hand carries the quiet uncertainty of whether the opponent is learning, adapting, and reacting—or simply executing code. Until detection meaningfully outpaces evasion, that uncertainty will remain part of the modern online game.
