DraftKings built AI to find losing bettors, shelved the addiction model
A New York Times investigation published Friday says the data DraftKings used to decide who deserved a free bet was never pointed at the customers those bets were most likely to hurt.
DraftKings built a machine-learning model in 2023 that scored customers on how much money they were likely to lose in response to a promotion, then used those scores to decide who got free bets and bonuses. The model was trained on customer betting records and ranked bettors by their projected response: the higher the score, the more money a person was expected to lose for every offer sent their way. According to the Times, several former employees who worked on the system say the company has only leaned further into it since.
The detail that carries the story is what the company did with the other model. Former DraftKings data analyst Jayden Butts told the Times the selection logic was explicit: "We are looking for traits and features that we can target that indicate a good investment." He added the quiet part out loud — the best investment would be a problem gambler. Meanwhile, a parallel effort to use similar prediction techniques to flag customers at risk of addiction was never finished, the reporting says. Same data, same team, two models; only the one that found losers shipped.
DraftKings disputes the framing. Its chief responsible gaming officer, Lori Kalani, told the Times the company monitors customers for risky behavior, and said it declined to deploy risk-prediction technology because it was not sufficiently evidence-based — an argument that cuts both ways, since the marketing model was deployed without anything approaching that standard of proof. The company said it "rejects any implication" that its promotions unfairly target customers, and maintains the offers go to users with sustained, engaged platform use. The Times reporting does not settle whether "engaged users" and "users predicted to lose" are in practice the same population; that depends entirely on how the model weighs its variables, which DraftKings has not published.
That gap is the real story. This is not a rogue deployment or an alignment failure — it is an ordinary commercial ranking model doing exactly what it was built to do, pointed at a population where the harm and the revenue are the same people. Every consumer-facing ML team makes this trade implicitly: the label you optimize is the policy you ship. DraftKings had the capability to build the protective model and the data to build it; what it lacked was a reason to prefer it.
The regulatory angle is why this lands now. Gambling regulators in the US have spent two years asking operators to demonstrate that their responsible-gaming tools are evidence-based, and the industry has answered by pointing at exactly the kind of risk-prediction system the Times says DraftKings left on the shelf. If the reporting holds, lawmakers get something more specific than general industry concern: a documented case of a company that built the tool, tested the tool, and chose not to use it.
What to watch: whether state regulators in the markets where DraftKings operates open a review, and whether the company publishes any detail on how its promotion model is constrained — the model card that does not exist is now the load-bearing question.
If a bonus offer landed in your inbox and a model had decided you were likely to lose, would you want to be told? Tell us in the comments.
