DraftKings built a model to target high-loss bettors

The model's target

DraftKings began developing the promotional model in mid-2023. Its task was not simply to predict who might place a bet, but whether a particular customer would generate more revenue than the company spent on incentives.

A high score meant that a player was likely to lose more money on each promotion and continue gambling. The company therefore focused on customers who appeared most likely to accept offers such as:

Sending those promotions to random customers would have been expensive. The model was intended to find people with stronger signs of gambling addiction, because they were more likely to accept the offers and keep playing.

Jayden Butts, a former DraftKings data analyst who tested the model, told The New York Times that the company was looking for signs of a financially valuable customer. From a purely financial perspective, he said, the most valuable customer was a person with a gambling addiction.

A former DraftKings analyst who left the company in 2024 described the practice as predatory. The logic, he said, was simple: the more a customer lost, the more the company gave them to keep playing.

What DraftKings tracked

The model drew on detailed records of user behavior. DraftKings analyzed:

The company also used another model to predict whether a customer was likely to stop playing. One employee said he worked on a system for identifying users ready to leave the platform and bringing them back.

The distinction matters. The same behavioral data could support two very different goals: finding customers who were likely to lose more, or finding customers whose losses suggested that intervention was needed.

The safety system that went nowhere

Several employees tried to build the second system. Nestor Hernandez began developing a model that would identify vulnerable players and assign them a risk score. DraftKings could have used it to warn a customer or block the account before losses became unmanageable.

Jake Shannin continued working on the project until it was stopped. The team was supposed to present the model to Lori Kalani, DraftKings’ head of responsible gaming, but the meeting was abruptly canceled on the day of the presentation.

The project was closed in 2025, several months after Hernandez left the company.

DraftKings told The New York Times that its marketing practices were not unfair or improperly targeted at customers. Kalani said the company’s leadership had collectively decided not to use predictive models to identify gambling addiction because the technology did not have sufficient evidentiary support.

That explanation leaves the central asymmetry intact: the company had enough confidence to model which customers could produce more revenue from promotions, but not enough confidence to deploy a model aimed at protecting vulnerable players.

My read is that the missing evidence is not just a technical detail. The announcement is quiet about what standard the revenue model had to meet before it could influence marketing, and why the protection model was stopped rather than tested under safeguards. What I’d want to know is whether “insufficient evidence” applied equally to both uses of prediction, or only to the use that might have reduced play.

The case shows the risk of treating behavioral prediction as neutral infrastructure. A system built from the same customer history can either limit harm or increase the value extracted from it. At DraftKings, the commercial use reached the point of targeted promotions; the protective use did not.