Times: DraftKings' Model Ranked Customers by Expected Losses — While Addiction Prediction Was Reportedly Stalled
A machine learning model built in 2023 scored the company's customers by how much money they were likely to lose per offer — while, according to former employees, similar technology was held back from predicting gambling harm.
Times: DraftKings' Model Ranked Customers by Expected Losses — While Addiction Prediction Was Reportedly Stalled
A machine learning model built in 2023 scored the company's customers by how much money they were likely to lose per offer — while, according to former employees, similar technology was held back from predicting gambling harm.
In 2023, DraftKings built a machine learning model based on customers' betting data to answer a concrete question: who was most likely to respond to promotions by betting — and losing — more? According to an investigation by The New York Times, published September 19, 2026, the model scored each customer by how much money they were likely to lose per offer presented to them.
The investigation, bylined Alex Klavens, Walt Bogdanich and Jenny Vrentas, is based, according to the newspaper, on internal documents and interviews with dozens of former DraftKings employees. Many of them were, according to the Times, concerned that the company's efforts were harming problem gamblers. It is worth being clear up front: all the details below come via the Times' reporting on internal material — AIMag does not have access to primary documentation from DraftKings, and the company's response to the investigation is not available in the material this article is based on.
Why the Model Was Built
DraftKings was, according to the Times, spending hundreds of millions of dollars every year on promotional incentives: "free" betting money advertised via emails and alerts on customers' phones. But the company knew little about how effective these offers actually were.
This is the problem the machine learning model was meant to solve. In practice, it comes down to pattern recognition: the model was built on customers' betting histories to identify traits of players who have historically responded to offers with increased betting and increased losses. The result was a per-customer score for expected loss per offer.
The Analyst Who Tested the Model
The most concrete account of how the system was used comes from Jayden Butts, a former data analyst at DraftKings who was named and quoted by the Times. His job was to test the model by prioritizing free bets and bonuses for the likely losers.
According to the Times, a question began to gnaw at him: weren't many of these same people vulnerable to addiction?
"We are looking for traits and features that we can target that indicate a good investment," he told the newspaper. By strict financial logic, he concluded, "the best investment would be a problem gambler." (Quoted from the Times' English original.)
The Asymmetry: Same Data, Opposite Uses
The core of the investigation is not that the company built a predictive model, but the reported imbalance in what the same data was used for.
According to six former employees who worked on the systems, DraftKings has continued to hone its methods for targeting likely-losing players with promotions that encourage more betting, following Butts' tests. At the same time, according to four other former employees, the company slowed or stopped attempts to use similar technology to predict who might develop a gambling problem based on their betting activity.
Methodologically, these are the weakest claims in the story. Both groups are anonymized former employees, and the characterization of what was "slowed or stopped" — and why — is the Times' rendering of their accounts. There is no DraftKings documentation in the available material that confirms or disproves them.
The Timing
The investigation was published on September 19, 2026, at the start of the NFL season's betting period — the stretch when promotional intensity in the industry is normally at its peak. The timing alone makes the story consequential: reporting on how the company's promotional machinery works internally lands precisely when the campaigns are most visible to customers.
What Remains Open
Several substantial questions cannot be answered on the available basis. First: DraftKings' response is not available in the material AIMag has had access to, and cannot be reproduced until it is verified. Second, the contents of the internal documents are known only through the Times' rendering; no primary documentation is available. Third, it is unclear exactly when the model went into full operation beyond the 2023 build-and-test period that Butts describes.
That leaves a story with an indirect but concrete evidence base: a named former employee describing the mechanism and his own conclusions, and groups of anonymized former employees whose accounts of halted addiction prediction are the Times' reporting, not verified facts. The same data stream could, in principle, be used to find the customers who lose the most — and to find players before harm occurs. Which of these goals DraftKings prioritized is, for now, a question only the company itself can answer.