The Washington Post used AI to analyze 50 televised sports games for references to betting. Here’s how we did it.
How we picked what sports to watch
To model the experience of a casual sports fan, we analyzed games aired on cable or broadcast TV (except two, from the NBA TV streaming service) in Washington between Dec. 21, 2025, and Jan. 23, 2026. League schedules constrained our choices: NCAA football was in its bowl season; the WNBA and MLB were not included because they don’t play in winter.
We selected games to maximize variance by sport, league, gender and channel, and chose a wide variety of teams and stadium or arena locations. Other factors also can potentially affect gambling content — such as the legality of sports betting in the home states of the teams playing — but were not controlled for in this analysis.
We chose 16 hours of NFL content, six hours of college football, eight hours of hockey, eight hours of professional men’s basketball, six hours of men’s college basketball and six hours of women’s college basketball.
An hour of each game was chosen randomly, including commercials and other content, starting no earlier than kickoff, puck drop or tip-off.
How we defined ‘gambling references’
We counted “references to gambling” as the names or logos of sports betting apps, state lotteries or casinos. We also counted commentators’ references to gambling lines or wagers, and odds or over/under lines displayed on sports tickers or on-screen graphics.
We didn’t include gambling metaphors from commentators (“They’re playing with house money”), a car company’s betting-themed sweepstakes or ads for an investment app that offers prediction market bets through a partnership with Kalshi.
How we refined our prompt
An internal Post video tool called Haystacker extracted a still frame from the videos every two seconds and asked an AI model to examine each one, using a custom prompt meant to focus on gambling-related imagery. We combined that with transcribed audio from the same games.
We refined that prompt over several iterations, to make sure Haystacker didn’t misidentify non-gambling-related things (false positives, measured as precision) or miss them (false negatives, measured as recall). We pitted its results against a hand-labeled evaluation set, created by three Post reporters who watched 214 random 10-second chunks of football and basketball footage.
Our final Haystacker prompt, which used OpenAI’s GPT-5, got a precision rate of 88 percent; 21 of the 24 snapshots flagged as gambling references really were. The other three were false-positives. The recall was 100 percent; the model found all the examples of gambling that Post journalists identified. (Our hand-labeling wasn’t perfect, either: GPT-5 found a few subtle gambling references we humans missed.)
The Haystacker pipeline could potentially miss some gambling references that appear on-screen for less than two seconds — but we found no instances of this in our evaluation set.
After examining Haystacker’s results, we added a second step to categorize gambling references by type, using a taxonomy modeled on work by Raffaelo Rossi, a U.K. marketing professor who has done similar analyses. We used GPT-5 to classify each gambling-reference frame as containing a commercial, broadcast overlay, ticker, sponsored segment, field or stadium signage, jersey sponsorship, network studio branding, or a responsible gambling message. (Each frame could contain multiple.) It also extracted the name of the gambling company referenced.
This second stage had a precision of 97 percent and a recall of 98 percent for classification and a precision of 98 percent and recall of 97 percent for identifying brands. The accuracy rate of our final results is somewhat higher, because we hand-corrected errors we found during the reporting process.
Read the full story about what we found here.
The post Here’s how we used AI to find gambling ads in televised sports appeared first on Washington Post.
