The Analyst Who Went Looking For Good Investments
Jayden Butts spent part of 2023 building a scoring model at DraftKings that ranked bettors the way a lender ranks loan applicants, except the trait his model was optimizing for was continued losing. The model combined play frequency, account balance and loss-to-wager ratio into a single number, then predicted how much extra a customer would wager if DraftKings sent a promotional offer. Employees called the metric elasticity.
"We are looking for traits and features that we can target that indicate a good investment," Butts told the New York Times, whose investigation drew on more than 40 former employees, internal memos, Slack messages and betting records. One unnamed former employee put the logic more bluntly: the best investment would be a problem gambler.
The Detector That Never Got The Same Attention
DraftKings also built a second model, one meant to flag customers showing signs of problem gambling before they lost money the company could not ethically keep. Four former employees told the Times that detector stalled or was shut down outright, even as the elasticity model kept being refined through 2025.
DraftKings disputes the framing. The company said its promotions go to "customers who demonstrate sustained, engaged use of our platform, not toward customers based on their losses." It has not disputed that both models existed.
Two Signals, One Budget
DraftKings ran two AI models with opposite goals, and only one of them competed for revenue. A model built to find customers about to lose money the company should stop chasing does not produce a visible number on a balance sheet when it works, because success there means revenue never collected. A model built to find who will wager more after a promotion shows up directly in sportsbook margins. When an engineering team has to decide which model gets the next sprint, the model tied to revenue wins by default unless someone with authority decides, explicitly, to fund the other one anyway.
DraftKings credited part of a 13 percent improvement in promotion-driven sportsbook margins in 2025 to this targeting work. The problem-gambling detector never had an equivalent number, because there is no metric that rewards not making money from a customer.
Massachusetts Already Knew
The state did not discover this gap when the Times called for comment. The Massachusetts Gaming Commission had commissioned a study from UNLV's International Gaming Institute, released in November 2025, that identified a governance gap between what AI tools in gambling marketing could already do and what the commission's existing rules covered. That study led to an AI task force inside the commission, tasked with watching exactly this kind of practice.
That task force sat until the Times published. On Thursday, chairman Jordan Maynard said the commission shares "the concern" over how fast these technologies are moving, and commissioner Paul Brodeur called the report troubling, noting Massachusetts is one of the only states that references AI in its gambling rules at all. Executive director Dean Serpa has now been told to engage DraftKings directly, and the review will extend to every sportsbook the commission licenses, not DraftKings alone.
What This Means If You Run Two Models Like This
Nothing about this requires bad intent to explain it. A model tied to revenue will get resourced ahead of a model tied to harm prevention unless someone with authority makes an explicit, ongoing decision to fund the second one at the same priority as the first. That decision does not happen by default, and a company that skips it will end up exactly where DraftKings did: a working system for the profitable half of the problem and an idle one for the other half.
The same applies to the regulator's side of this story. Commissioning a study that names a governance gap is not the same as closing it; Massachusetts had its own diagnosis ten months before it acted on it. In Europe, a comparable scoring system used to rank people by financial risk would fall inside the EU AI Act's high-risk category, with mandatory human oversight built in from the start rather than added after a newspaper asks a question. Building the detector, or commissioning the study, is the easy half. Deciding, on purpose, to give it a budget and a deadline is the part that actually needed a decision.
Read next: One Missed Deadline Broke Two of Apple's Promises | EVE Online's Studio Hired an Ex-Minister to Run Its AI Program
