ESPN's AI Tells Detector at 2026 WSOP Faces Skepticism Over Training Data Limitations
An AI tool designed to read poker players' physical tells debuted at the World Series of Poker Main Event, but experts question its accuracy given its small dataset.
Last verified:
BLUF
ESPN introduced an AI tells-detection system during the 2026 World Series of Poker Main Event’s live broadcast in early July, designed by Air Force engineer Luke Geel to predict player hand strength from physical tells. However, professional poker players including 17-year veteran Michael Gagliano have raised serious doubts about the tool’s accuracy, arguing that it was trained on an insufficient dataset—camera feeds from only three broadcast tables covering a small fraction of the tournament’s 9,000+ entrants.
How the AI Tells Detector Works
According to Wired AI, the system ingests video signals tracking multiple behavioral metrics: eye movements, blink frequency, posture shifts, chip-handling patterns, and hand fidgeting. The model then correlates these observations with hand outcomes to assign a probability distribution across four hand categories—strong made hands, drawing hands, bluffs, and others—displayed as an overlay during broadcast moments.
The architecture mirrors human tells-reading, which professional poker specialists have practiced for decades to gain competitive advantage in a game of incomplete information. By automating tells detection, the tool ostensibly democratizes pattern recognition that previously required years of table experience.
Data Bottleneck and Expert Skepticism
The critical limitation emerged quickly. The 2026 WSOP Main Event drew over 9,000 entries, but the AI’s training set consisted solely of camera feeds from three broadcast-equipped tables—the same feeds ESPN used for live transmission. This means the vast majority of players never appeared in the training data, and those who did sat for only brief windows during the tournament’s two-and-a-half-week run.
According to Wired AI, Gagliano, who reached this year’s Main Event final table competing for the $10 million top prize, examined every second of ESPN’s broadcast streams during the event’s break. His conclusion: the sample size was too small to capture poker’s strategic variability. “The streams are varied enough that you don’t get the same players too frequently,” Gagliano told Wired, implying that the model lacked sufficient hand-history repetition per player to build statistically valid patterns.
Other poker professionals quoted by Wired AI echoed this skepticism, treating the tool as either entertainment or a premature application of machine learning to a domain requiring deeper data.
Why This Matters
If Geel’s tool succeeds only as broadcast theater, the immediate impact is negligible—ESPN gains a novelty segment, viewers get a talking point, and competitive poker proceeds unchanged. But the skepticism reflects a deeper tension: AI’s hunger for scale versus poker’s inherent scarcity of public, high-stakes data.
A fully realized tells detector—trained on years of archived broadcast footage from multiple operators, multiple tables, and diverse player populations—would threaten the information asymmetry that makes poker skill legible. Players who cannot control their physical behavior would face compounding disadvantage. That prospect has already begun shaping industry conversation, even if the current iteration lacks the training horsepower to validate such concerns. For now, Luke Geel’s system remains a proof-of-concept that arrived before its data.
Frequently Asked Questions
What exactly does ESPN's tells detection tool do?
The system analyzes video feeds to track player movements—eye contact, blink rate, posture, chip handling, hand fidgeting—and applies a hand-strength model to predict whether a player holds a strong made hand, a drawing hand, or is bluffing.
Why are poker professionals skeptical about the tool?
According to Wired AI, the AI was trained only on broadcasts from three camera-equipped tables during the 2026 WSOP Main Event, capturing fewer than 9,000 of the tournament's 9,000+ entries. Most players never appeared on those feeds, and even those who did lacked sufficient hand history for robust pattern training.
Could this tool reshape competitive poker?
Unlikely in the near term. The sample-size and diversity limitations mean the tool's predictions lack statistical backing. Longer-term, if trained on larger datasets, AI tells detection could shift the game's emphasis from physical deception to pure strategic play.