ESPN rolled out an AI tells detection tool during its broadcast of the 2026 World Series of Poker Main Event, overlaying live metrics on players' eye movements, blink rates, posture, and chip handling alongside a hand-strength probability chart. The tournament drew more than 9,000 entries competing for a $10 million top prize, but the AI only trained on footage from three filmed tables. The system was built by Luke Geel, an AI engineer at the US Air Force, and appeared periodically starting in early July before being pulled ahead of the final table.
The tool ingests player-behavior signals — eye movements, blink cadence, posture shifts, chip handling, hand fidgeting — and cross-references them against the outcomes of hands captured on camera. From that, it estimates whether a player is holding a strong made hand, a drawing hand, or a bluff. The approach mirrors what human tells specialists do, except a camera-based model sees only what fits in frame.
The core weakness is sample size. Of the 9,000-plus entries, only a small subset played at the three tables ESPN filmed, and few of those sat there long enough to generate a useful behavioral profile. Michael Gagliano, a 17-year professional who made the final table in eighth chip position, spent the two-and-a-half-week break combing every second of the live streams for reads on his remaining opponents.
Key facts
- 01ESPN deployed an AI tells detection tool during the 2026 World Series of Poker Main Event, which drew over 9,000 entries and a $10M top prize.
- 02The system, built by US Air Force AI engineer Luke Geel, tracks eye movement, blink rate, posture, and chip handling to predict hand strength.
- 03Training data was limited to hands captured at three filmed tables — a fraction of the tournament's total action.
- 04Omaha Productions confirmed the tool would not be used at the final table reached in mid-July, without giving a reason.
- 05Shaun Deeb, who finished 15th, called the feature a 'swing-and-a-miss' and said he'd bet his human tells team against the AI.
Even with weeks of footage, Gagliano said he came up mostly empty. The tool faces the same ceiling: without repeated exposure to the same player across many hands, a model cannot learn the individual quirks that make tell detection work. Gagliano is still competing this week for the $10 million.
Shaun Deeb, a two-time WSOP Player of the Year who finished 15th in the 2026 Main Event, pushed back harder on the framing. He referenced the Oreo-cookie tell scene from the 1998 film Rounders — Matt Damon's Mike McDermott folding a monster to John Malkovich's Teddy KGB — as a caricature of how tells actually work in practice.
Deeb's point is that a camera-based system misses most of the signal. Pulse, breathing rhythm, subtle leg movement, verbal cadence, and physical positioning are all live-read inputs that a broadcast feed either compresses or cuts entirely. A model watching TV footage is training on a lossy version of the actual game.
There is also the intention gap. An AI can identify that a player looks confident, but confidence about what — a made hand, a semi-bluff, a marginal call in a high-variance spot — is a judgment call that depends on the player, the situation, and the stakes. Gagliano noted that his own body language in a Main Event pot might read as nervous even with a strong hand, simply because the pressure is different from a casual game.
Geel, the tool's creator, has been open about the limits. He told Wired via email that a larger sample of hands would improve accuracy and that blind tests against other poker competitions produced mixed results. Omaha Productions, the company that licenses WSOP coverage for ESPN, confirmed the tool was pulled from the final table broadcast in mid-July without explaining why.
The longer-term question is whether tools like this get deployed away from broadcast, in the high-roller circuit where six-figure buy-ins bring the same recognizable pros to the felt dozens of times a year. Hundreds of hours of footage of top players already exist, and studying opponent streams is standard prep. A better-trained model with a bigger corpus is a plausible next step, though live tables ban electronic aids, and Deeb expects Meta smart glasses to face similar bans as they proliferate.
For AI Chat Daily, the interesting angle is not whether an AI can outread Shaun Deeb — it can't, yet — but the pattern of computer vision being sprinkled onto live broadcasts as a graphics layer rather than a decision tool. Sports coverage will keep testing these overlays because they are cheap to produce and fill airtime, and most viewers will not audit the underlying accuracy. The risk for AI credibility is when the on-screen confidence of a slick text overlay outruns what the model actually knows, and poker, a game built on hidden information and small samples, is a hard first test to pass.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
