What the model did
DraftKings built a machine-learning model in 2023 that scored customers by how much money they were expected to lose after receiving a free bet or a bonus, the New York Times reported on Saturday, citing interviews with former employees, as summarised by Yahoo News.
The inputs were ordinary: how often someone played, what their account balance looked like, the ratio of losses to wagers, and how likely they were to stop gambling altogether. The output was a single number staff referred to as an “elasticity” score. A higher score meant a customer was worth chasing with more perks.
Why it worked
The model was not built to find problem gamblers. It was built to find the best return on a promotional dollar, and the two turned out to point at the same people. Former employees told the Times that the financial logic identified problem gamblers as the most lucrative targets for bonus offers.
“We are looking for traits and features that we can target that indicate a good investment,” former data analyst Jayden Butts told the paper.
That sentence is the whole story in miniature. Nothing in it is about gambling harm. Nothing in it needs to be, for the harm to follow.
The model that did not ship
In mid-2024, DraftKings data scientist Nestor Hernandez began building a different model: a risk score meant to spot customers heading toward a gambling crisis days or weeks in advance, by scanning money movements, loss-chasing patterns, age and gender.
It ran on much the same infrastructure as the elasticity score. It was sidelined. The company’s responsible gaming officer said the risk-prediction tool lacked sufficient evidence-based support, according to the Times.
The asymmetry is the finding worth keeping. A predictive model aimed at revenue was accepted on its results; a predictive model aimed at harm was held to an evidentiary standard the first one was never asked to meet.
What DraftKings says
The company disputes that it targeted people by predicted losses, saying promotions go to engaged users, and points to its responsible-gaming partnerships and monitoring tools.
It has also told investors what the analytics were worth: DraftKings credited data science with improving promotion-driven sportsbook margins by roughly 13% in 2025, across promotional spending it describes in the hundreds of millions of dollars.
Why this matters beyond betting
Gambling regulators in the United States supervise advertising, bonus terms and self-exclusion. They do not generally supervise the ranking function that decides who sees an offer. That gap is not unique to sportsbooks; it is the same gap that exists wherever a model sorts customers by expected value and the sorting itself is invisible to the person being sorted.
Watch for whether any state regulator asks to see the feature list — the question that decides this is not whether AI was used, but whether “likelihood of stopping” is a legitimate input to a marketing model.
