The World Series of Poker has long turned uncertainty into entertainment. During the early broadcast days of its 2026 Main Event, ESPN added a new layer: an artificial intelligence system that tried to decide whether players held a strong hand, were drawing or were bluffing.
The “AI tells detection” tool was built by Luke Geel, an AI engineer for the US Air Force. It analyzed every hand captured on camera at the tournament and assembled a database of more than 9,000 entries. Its inputs included eye movement, blink rate, posture, chip handling and hand fidgeting. On television, that can look like a machine reading the people who make their living reading one another.
Yet the experiment revealed as much about the limits of behavioral AI as its possibilities. Poker professionals described a far messier contest, in which visible motion matters only when combined with history, context and signals the cameras may never capture.
A large database can still be a small sample
More than 9,000 observations sounds substantial, but only three tables were recorded. Most competitors therefore never entered the system, while many of those who did appeared for too little time to establish a reliable personal baseline. A model may detect that somebody blinked or moved chips differently; it still needs enough comparable situations to know whether that change means anything.
Michael Gagliano, a 17-year professional who reached the final table and was competing for a $10 million top prize, reviewed every second of the available streams during a two-and-a-half-week break. Even with that unusually focused preparation, he found the information of limited practical value.
Geel acknowledged that more observations would improve the model. He also reported mixed results from blind testing. Those qualifications matter because the tool’s labels can easily appear more certain than the evidence beneath them. A prediction is not a view of the hidden cards; it is an interpretation of incomplete behavior.
Poker tells depend on context the camera cannot see
Experienced players do study physical signals, but their accounts extend well beyond the features available to the system. They may consider leg movement, breathing, pulse, verbal behavior and how someone checks their cards. More importantly, a movement does not carry a universal meaning. The same gesture may reflect strength, stress, fatigue or simply a player’s normal routine.
That creates two linked problems. First, cameras capture only a restricted angle, excluding signals below the table or outside the frame. Second, even a correctly measured change does not directly disclose intention. Computer vision can convert motion into data, but the inference from motion to bluff still depends on situation and individual history.
The gap is especially important in poker because participants actively manage the impressions they create. The system is observing opponents who know that observation is part of the game. A model trained on visible behavior may eventually influence that behavior, making its own data less stable.
The near-term value may belong to the broadcast
An Omaha Productions representative said the tool would not be used during the final table, without explaining why. Its absence from the event’s decisive stage underlines the difference between an engaging broadcast graphic and dependable competitive intelligence.
Players also cannot simply run such software beside them. Tools are not allowed at live tables, and some events prohibit electronics. Professionals interviewed by Wired still favored in-person specialists who can observe the whole player and interpret behavior within the flow of the game.
The longer-term opportunity is not imaginary. High-roller tournaments repeatedly feature the same professionals and can produce hundreds or thousands of hours of footage. That richer history could support stronger individualized comparisons, although it would not eliminate hidden signals or ambiguous human behavior.
For now, the system works best as an experiment in making probability visible. It gives viewers another way to engage with a hand and offers researchers a demanding test of computer vision. What it does not provide is an oracle. Poker’s tension survives because a measured tell remains evidence to interpret, not an answer.
