DraftKings built a machine-learning model in 2023 to work out which customers would respond to a free bet or a deposit bonus by gambling — and losing — more money, then used it to decide where its promotional spending went. A second model, built on the same betting records by a data scientist inside the company, would have scored customers for the risk of a gambling crisis and let DraftKings intervene before one arrived. The first ran on live customers. The second was shut down before it was ever presented.
That is the shape of the New York Times investigation published 19 September, based on internal research memos, presentations, Slack messages and betting records from experiments run on customers, plus interviews with more than 40 former DraftKings employees.
The model that shipped
DraftKings was spending hundreds of millions of dollars a year on promotional incentives — “free” betting money pushed out by email and phone alert — and, according to one internal memo, had “no information or insight on expected user-level profitability of reinvestment dollars.” It did not know which offers produced losses.
In mid-2023 the casino division set out to fix that with a machine-learning model built on customer betting records. Jayden Butts, a data analyst who joined in 2022, was assigned to test it on real customers. The score it produced was known internally as “elasticity.” The question it answered, as Butts described it: “Is this person going to give us more than we’re giving them? And if the answer is yes, open the floodgates.”
The inputs were behavioural: how frequently someone played, their daily account balances, and how much they typically lost compared with how much they bet. The model also incorporated a second model that calculated how likely a user was to stop gambling. Users scoring below average were marked “inelastic” and got fewer incentives. Butts began testing on roughly 5,000 casino players in September 2023, then widened it. He told the Times he had assumed the point was to avoid unprofitable customers, and was told the aim was not to cut spending but to “redeploy” it — toward the biggest losers.
Eight former employees told the Times that similar models for sports betting promotions followed, and that the techniques have continued to be developed this year, including one system to identify users at risk of leaving the platform and coax them back, and another that ramps up promotions for people gambling more aggressively than usual. Two analysts said they independently raised the concern that these models would identify vulnerable gamblers as targets, and asked for protections. Both were told no, they said — that would be handled by another part of the company.
A 2023 internal memo reported that slots revenue was “more elastic” than revenue from other casino games, meaning promotions were particularly effective at driving users to them; a Times analysis of internal casino betting data from early 2024 found that high-elasticity bettors gambled more on slots on average than low-scoring ones.
DraftKings rejected the implication that its marketing “improperly targets customers,” saying promotions are “directed toward customers who demonstrate sustained, engaged use of our platform, not toward customers based on their losses.” It called Butts’s test “preliminary and inconclusive” and said it could not give a “definitive response” to his account, the memo or the Times’s analysis because it had not seen or verified the material. It disputed his account of his dismissal — he was fired in late 2024 for performance reasons, he said — but did not elaborate.
The model that did not
Nestor Hernandez, a computer science PhD, started building the mirror image in mid-2024: a system for the responsible gaming division that analysed the same betting records — deposits and withdrawals, attempts to chase losses, age and gender — to anticipate a customer’s impending crisis. “The idea of this model is to be more proactive instead of being reactive,” he said. “You will basically predict that a user will be in trouble, let’s say, a few days or a few weeks in advance. And you can act accordingly.”
He left in November with the model unfinished. A new team continued the work and concluded, in the words of member Jake Shannin, that “it is possible even with simple rules to get a better-than-guessing sense” of which customers would need intervention. In early 2025 the team prepared to present it to company officials, including chief responsible gaming officer Lori Kalani. The meeting was cancelled the day it was scheduled. Two other attempts to build similar algorithms were also shelved, per two former employees.
Kalani said leaders made a “collective decision” not to use predictive technology for problem gambling: “We evaluated that it wasn’t evidence-based.” She said DraftKings monitors customers for “potentially risky behaviours” and follows New Jersey regulators’ guidance on automatic triggers — more than two dozen indicators nationwide — escalating from a responsible-gaming message to an educational video to a questionnaire, and sometimes closing an account. The company declined to share the thresholds that prompt a response.
Several third-party firms sell versions of risk scoring, and other operators say they use them, including FanDuel and Fanatics. Phil Sherwood, senior director of responsible gaming at PrizePicks, said his company had a “moral obligation” to use such tools: “Frankly, without technology like this, I can’t look you in the eye and say, ‘We’re doing our best.’” DraftKings does use large language models for one responsible-gaming purpose — scanning communications from customers to company representatives for signs of distress. As the Times notes, the technology cannot flag customers who say nothing.
The arithmetic that made the choice
DraftKings took in around $8.7 billion in gross revenue from sports and casino gamblers last year and gave out about $3 billion in promotions, according to research by Citizens Bank. An executive recently told investors that data science and analytics improved its margins on promotion-driven sports bets by 13 percent in 2025, and that the company used A.I. to personalize hundreds of millions of dollars of promotional spending.
Its online casino games — the division where the elasticity model was built — bring in nearly a third of company revenue while being offered in only a handful of jurisdictions, and slots are widely regarded by public health experts as among the most addictive forms of gambling. The company reports 11 million customers, up from 5 million in 2022, across a business that has expanded into online casino and, more recently, prediction markets.
Responsible gaming runs through a division of around 50 people at a 5,000-person company. One former employee described it as “at odds with the rest of the company,” because its function is to slow gamblers down and slowing them down makes them less profitable. DraftKings said there was “no tension” between its business goals and its responsible gaming work.
The people on the other end
Bryan Biehl, of West Chester, Pa., lost nearly $70,000 gambling online, more than half of it at DraftKings. In late 2024, when he started therapy for his addiction, his inbox became the risk: “I would get flooded with bonuses and deposits. If you are in addiction, you are not going to say no.” Emails show 40 promotions from DraftKings in the first two weeks of December 2024. He bet once more on Christmas Day, then put himself on self-exclusion lists that blocked him from gambling apps. DraftKings declined to comment on his account.
Hassel Velasco, who lasted five months on the customer service team in 2024, told the Times he heard complaints about a dozen times from problem gamblers who kept receiving promotions. “It pushes people to almost chase their losses,” he said. Jacob Shulkin, a former data scientist who worked on promotions, described the mechanism from the inside: “I feel I’m getting free money, but really it’s dragging me back in.”
Asked whether promotions contribute to problem gambling, Kalani said they are “a marketing tool that every company out there uses,” comparing DraftKings to Amazon. “Shopping can be problematic for people.”
What the same data would have supported
The investigation does not allege a rule was broken. Its finding is narrower and harder to dismiss: the company held the data needed to score customers for risk and built the version that converted them into revenue, while the version that would have identified people in trouble was cancelled, twice more attempted and shelved, and left to competitors to use.
The public-health cost of that choice is booked outside the company. Calls about sports betting to Ohio’s problem gambling help line have more than quadrupled since the state legalized it in 2023; the state’s outgoing governor, Mike DeWine, told the Times that signing the law was his biggest regret. A study published last year in JAMA Internal Medicine traced a rise in Google searches like “Am I a gambling addict?”
Existing oversight runs on the reactive version — deposit thresholds and behaviour triggers that a customer has to trip first. The proactive version, on the evidence in this article, is technically available, cheap enough that smaller rivals license it, and available to the operator that declined it.