Security1 publisher2 min readPublished
DraftKings model scored gamblers on how much they would lose after promotions, the Times reports
DraftKings scored customers on how much they were likely to lose after a promotion, the New York Times reported after interviewing over 40 former staff. A ProPublica reporter betting his own money saw the offers grow as his losses did.
The Watch · Security desk
What happened
- The model drew on each customer's betting frequency, losses and daily account balances, according to the Times.
- After ProPublica's Jake Pearson lost nearly $1,800 on basketball in one evening, DraftKings invited him to a VIP tryout, and full membership followed later.
- DraftKings occasionally sent Pearson automated reminders to bet responsibly, and he found they were no match for a product design pushing him to gamble more.
- Employees pushed for a model to flag customers heading toward crisis; data scientists said it showed promise, but DraftKings shut it down after canceling a demo.
- DraftKings denied unfairly targeting gamblers and said its promotions reward engagement rather than losses.
Compiled by The WatchSomething wrong?How this is made
Why it matters
- contradiction DraftKings' engagement defense and the Times' account of a score built on expected post-offer losses describe different targeting rules, and former staff have already described the second one to reporters.
- exposure With a promising harm-detection model on record as shut down, DraftKings will have a harder time arguing it had no way to spot customers heading for crisis.
- exposure Online casino games, which DraftKings now runs in select jurisdictions, put more products within reach of the same kind of loss-based customer scoring.
According to the Times, the newer system dug through that data on its own and estimated how much each customer was likely to lose after receiving a promotion [4]. It turned the estimate into an "elasticity" score and used it to pick out customers likely to lose more in response to offers [4]. The offers were promotions such as profit boosts on winnings, sent to bring people back to the site [3]. Former insiders told the paper that development began in 2023 [7]. The model replaced a cruder system that crunched basic numbers to report on a user's behavior [7].
Jayden Butts, a former employee who helped develop the system, told the Times it would help the company maximize the return on its promotional spending [5]. In a sportsbook, the return the score was built to maximize is the customer's expected loss after taking the offer [1].
Beyond the interviews, the Times reviewed Slack communications, internal memos and betting records [6]. ProPublica's test was a single account. Jake Pearson spent thousands of dollars betting on the app to see how its marketing would treat him [10].
The at-risk model was not the only attempt. Data scientists told the Times that later efforts to build similar systems inside DraftKings also failed [9].
The World Health Organization estimates that 1.2% of adults worldwide have a gambling disorder, and that people betting at harmful levels generate around 60% of gambling losses [16]. Those are global estimates. On those proportions, a score that ranks customers by expected loss after an offer would put harmful-level bettors high on its list [2]. The Electronic Frontier Foundation has said AI "supercharges the harms of online behavioral advertising" [15].
The account here relies on Malwarebytes' summary of both investigations [18]. That summary does not describe any regulator response.
What to watch
- Whether a state gaming regulator asks DraftKings for the elasticity model's design or for the memos and Slack records the Times reviewed.
- Whether DraftKings revives a harm-detection model or publishes how it decides which customers receive promotions.
- Whether further reporting finds the same loss-response scoring applied to DraftKings' online casino products.