ESPN’s use of an AI tells detection tool during the 2026 World Series of Poker Main Event put a familiar poker question into a new frame: can software read a player’s body language well enough to say something useful about a hand?
The feature appeared during portions of the live broadcast in early July. It showed movement-based metrics and a “hand strength model” chart that suggested broad possibilities, such as whether a player might have a strong made hand, a drawing hand, or a bluff.
For viewers, the appeal is obvious. Poker is a game built around hidden information, and televised poker has long tried to make that hidden layer visible. But for players, the tool raised a harder question: whether AI can turn a small set of camera feeds into meaningful insight about high-stakes human behavior.
What the AI tool was trying to read
The system was designed by Luke Geel, an AI engineer for the US Air Force. According to the source article, it watched every hand captured on camera in the 2026 WSOP Main Event to build a tells database on players who appeared on the broadcast feeds.
Its inputs included visible details such as eye movements, blink rate, posture, chip handling movements, and “hand fidget” metrics. It then compared that information with hand outcomes to estimate the likelihood of different general hand types.
That goal mirrors a long-standing part of serious poker. Skilled players often look for conscious movements, body language, and subconscious tics that may reveal something about an opponent’s plan. The difference is that ESPN’s feature tried to turn that observational work into a broadcast graphic.
The result was a polished on-screen layer, but its polish also made its limits important. A chart can look precise even when the underlying read is uncertain. In poker, that distinction matters because confidence, nerves, and hand strength do not always line up neatly.
Why pros questioned the sample size
The central concern from poker experts in the source article was data. The 2026 edition of the WSOP Main Event drew over 9,000 entries, but most players were never seated at one of the three tables recorded by cameras.
That means the same camera feeds used for the broadcast were also the pool of footage available to train the AI tool. Even players who did reach those featured tables were not necessarily shown often enough for a deep behavioral profile.
Michael Gagliano, a 17-year poker professional who made the Main Event final table and was playing for the $10 million top prize, described the problem directly: “The streams are varied enough that you don't get the same players too frequently.”
Gagliano started the final in eighth chip position. During the two-and-a-half-week break after the final table was reached in mid-July, he reviewed every second of ESPN’s live streams to search for useful tells or information on the remaining opponents.
Even with that level of effort, he said the footage was limited. “I don't know how much actual information I'm going to be able to act on from what I saw,” he said.
That point applies to AI as well. A model can only learn from the situations it has seen. If a player appears briefly, or if the recorded hands do not cover enough different pressure points, bet patterns, and outcomes, the system may be trying to infer too much from too little.
Poker tells are not just visible gestures
The debate was also about what counts as a tell. A camera-based tool can track visible and audible patterns, but poker professionals argued that meaningful live tells are broader and more situational than a simple movement reading.
Shaun Deeb, a two-time winner of the WSOP Player of the Year award, said the popular image of tells is too narrow. Deeb, who finished 15th in the 2026 Main Event, pointed to the well-known Oreo cookie tell from the film Rounders as an example of how the public may think about the subject.
“Physical tells are so much more expansive than I think the public realizes,” Deeb said. “There are leg tells, checking tells, verbal tells, breathing tells, pulse tells. There's an insane amount of tells available, and most of those can't be picked up by a camera.”
That limitation matters because a visible signal does not automatically reveal intent. A player may appear confident because of a strong hand, because they misunderstand the situation, or because the stakes change how they behave.
Gagliano gave a plain example around two pair. The same hand can feel powerful to one player in one setting and fragile to another player in a tougher spot. In the Main Event, body language might reflect the pressure of the situation rather than the objective strength of the cards.
A broadcast feature, not an oracle
For television, the standard may be different from the standard at the table. A broadcast feature can be useful if it gives viewers another way to think about a hand, even if it is not accurate enough to guide a professional decision.
The source article makes clear that Geel did not present the system as perfect. He said a larger sample of hands would be better and told WIRED by email that blind tests on other poker competitions had produced mixed results.
Some players remained unconvinced that the feature helped the broadcast. Deeb said, “I think they randomly found something to try to make it like another sport, and I just think it was swing-and-a-miss.”
The tool also did not continue all the way through the event. A representative from Omaha Productions, a company licensed by ESPN for WSOP and other sports coverage, said in a text message that the tool would not be used for the final table. The representative did not provide a reason for that decision.
What this could mean for poker footage
The larger issue is not whether one early broadcast tool was flawless. It is whether AI could eventually improve a process poker players already use: studying streamed and broadcast footage to learn about regular opponents.
The source article notes that in the “high-roller” tournament scene, buy-ins frequently reach six figures, and a smaller group of recognizable professionals often play each other in events that are broadcast. For those players, there may be extensive footage across many appearances.
That creates a different kind of data problem. The WSOP Main Event broadcast may not show enough of most players, but repeated coverage of the same professionals could give future systems more material to analyze.
Still, the same basic caution remains. Poker tells are not isolated movements with fixed meanings. They are shaped by context, stakes, player skill, table dynamics, and the gap between what a player believes and what is actually true.
ESPN’s AI poker tells tool therefore sits between entertainment and analysis. It can make a hidden part of poker feel more visible to viewers, but the professionals in the source article were clear about the risk: when the data is thin and the behavior is complex, a confident-looking AI read may be less decisive than it appears.