NYT: DraftKings Scored Customers by Expected Losses and Shelved a Harm Model

A 2023 model ranked online casino players by how much they would lose per promotion and steered bonuses towards the top scorers, according to the New York Times. A parallel model to flag customers sliding towards harm was shut down in early 2025. DraftKings says predictive risk scoring was not evidence-based.

• DraftKings built a machine-learning model in 2023 that assigned each online casino customer an "elasticity" score from playing frequency, daily balances and the ratio of losses to wagers, with a higher score meaning more expected loss per promotion, according to a New York Times investigation published on Friday
• The Times reviewed internal memos, presentations, Slack messages and betting records and interviewed more than 40 former employees; six said the company has continued refining the methods to direct promotions towards losing gamblers
• A separate model begun in mid-2024 by data scientist Nestor Hernandez to flag customers approaching harm was left unfinished when he left in November 2024, and a planned presentation was cancelled and the project shut down in early 2025, four former employees told the Times
• Chief responsible gaming officer Lori Kalani said leaders made a "collective decision" against predictive tools because the approach was not "evidence-based"; DraftKings said promotions go to customers with "sustained, engaged use" and called the 2023 test "preliminary and inconclusive"
• DraftKings gave out about $3 billion in promotions last year against roughly $8.7 billion in gross gambling revenue, the Times reported citing Citizens research, as lawmakers press FanDuel over its VIP programme

One Model Went Into Production and the Other Went Into a Drawer

DraftKings used machine learning to identify the customers most likely to respond to a free bet or bonus by betting, and losing, more, and directed promotional spending accordingly, according to a New York Times investigation published on Friday 19 September. The reporting draws on internal memos, company presentations, Slack messages and betting records, and on interviews with more than 40 former employees, several of whom told the paper they worried the targeting was harming problem gamblers.

The model was built in 2023, at a time when the company was spending hundreds of millions of dollars a year on incentives without a clear picture of which offers worked, the Times reported. It assigned each online casino customer a score, which staff called an elasticity score, from inputs including how often they played, how their balance moved, how much they bet and lost, and how likely they were to stop. The higher the score, the more the customer was expected to lose for every promotion sent. Jayden Butts, a data analyst assigned to test it about a year into his job, prioritised free bets and bonuses for the highest scorers across thousands of casino players. "We are looking for traits and features that we can target that indicate a good investment," Butts told the Times. "The best investment would be a problem gambler."

Six former employees said DraftKings has since kept refining its data science to direct promotions to losing gamblers in ways that encourage more betting, and the Times reported the company later developed similar techniques for sportsbook customers. A DraftKings executive told investors that data science and analytics improved margins on promotion-driven sports bets by 13% in 2025 and that the company used AI to personalise hundreds of millions of promotional dollars, the paper noted. DraftKings says it has 11 million customers, up from five million in 2022.

The same data could have been pointed the other way, and briefly was. Nestor Hernandez, a data scientist, began work in mid-2024 on a model for the responsible gaming team that would use betting records, drawn from the same pool used for promotional targeting, to assign risk scores before a customer reached a crisis point. He left in November 2024 with it unfinished; another team carried it forward; by early 2025 a planned presentation to company officials was cancelled and the project shut down, four former employees told the Times.

DraftKings disputes the framing. Lori Kalani, its chief responsible gaming officer, told the Times that leaders reached a "collective decision" not to use predictive technology for problem gambling because the approach was not "evidence-based", and that the company's existing monitoring system was more appropriate. The company said it rejects any implication that its marketing is unfair or improperly targets customers, that promotions are aimed at customers with "sustained, engaged use" rather than at people singled out for losing, and that Butts's test was "preliminary and inconclusive". Asked by Fortune in 2024 whether AI could make betting more addictive, chief executive Jason Robins said the company was not using it that way and that AI could help identify risky behaviour.

The Times put the scale in context: DraftKings handed out about $3 billion in promotions last year against roughly $8.7 billion in gross gambling revenue, citing Citizens research, and Ohio's problem gambling helpline has seen calls about sports betting more than quadruple since legalisation in 2023. In Washington, the SAFE Bet Act introduced by Senator Richard Blumenthal and Representative Paul Tonko would bar sportsbooks from using AI to track betting habits and serve personalised offers.

The Inputs Are the Story, and They Are the Same Inputs a Harm Model Uses

Frequency of play, balance movement, loss ratio and churn risk are not exotic variables. They are the features every responsible-gambling model in Europe is built on, and the reason regulators there require operators to act on them. What the Times describes is a company that computed those features for every casino customer, used them to decide who received more money to gamble with, and declined to use the same computation to decide who should receive an intervention. "Not evidence-based" is a real objection to predictive harm scoring; the academic literature is contested and false positives carry costs. It is a harder objection to sustain when the identical model, trained on the identical data, is considered evidence enough to allocate hundreds of millions of dollars.

The Regulatory Reading Writes Itself, in Two Jurisdictions at Once

In the United Kingdom this would be a licence condition question: the Gambling Commission's customer interaction rules already require operators to use the data they hold to identify harm, and the player-protection models it expects are the model DraftKings shelved. In the United States there is no equivalent federal rule, which is why the story is landing as a political one. Lawmakers have already told FanDuel its answers on VIP hosting were misleading and incomplete, New Jersey's panel has backed curbs on ads aimed at problem gamblers, and the SAFE Bet Act's sponsors now have a documented example of the practice their bill would prohibit. State regulators who licensed DraftKings on the promise that regulated betting is safer than offshore betting have a concrete question to put to it.

The Defence Is Narrower Than the Denial

DraftKings's statement rejects an implication and describes an intent, "sustained, engaged use", without disputing that a loss-based score existed, was tested and was followed by continued refinement of the methods. A test can be preliminary and a model can still be in use two years later; the former employees say it is. The evidence-based objection to the harm model is the company's strongest ground, and it would be stronger still if the same standard had been applied to the promotional one. Until DraftKings publishes what its monitoring system actually does, the shelved model and the deployed one will be judged together.

Every operator holds this data. The Times has documented what one of the largest chose to do with it, and what it chose not to.