The Best Investment Would Be a Problem Gambler: Inside DraftKings' AI Targeting Machine

On September 19, The New York Times published the kind of investigation the AI industry spends a lot of time talking about in the abstract and very little time seeing in the wild: a concrete, well-sourced case of a large consumer company building a machine-learning model whose explicit purpose was to identify the customers most likely to lose money — and then aiming its promotional budget squarely at them.

The company is DraftKings, the Boston-based sportsbook and daily fantasy giant. The model, built in 2023 from the company’s own customer betting records, scored every user on a single question: who is most likely to respond to a promotion by gambling more — and losing more? The higher a customer’s score, the more revenue DraftKings could expect per promotional dollar spent. Free bets, bonus credits, and personalized offers then flowed toward the top of that distribution.

What the model actually measured

According to the Times’ reporting, corroborated by six former employees, the model ingested a set of behavioral features that will sound familiar to anyone who has worked in growth marketing or credit risk:

• Play frequency — how often a customer bets, and whether cadence is rising
• Account balances — how much money is sitting in, or flowing through, the account
• Loss-to-wager ratios — how much of what a customer stakes comes back as winnings
• Churn probability — a separate estimate of whether a user was likely to stop gambling altogether

Each of these features is individually defensible as a marketing metric. Together, they compose something sharper: a predictive profile of vulnerability. A customer who bets frequently, loses a high share of what they wager, keeps reloading their balance, and shows no sign of quitting is — in the model’s terms — a high-scoring account. In human terms, that profile describes someone drifting toward problem gambling.

The starkest quote in the investigation comes from Jayden Butts, a former DraftKings data analyst who worked with the data: “We are looking for traits and features that we can target that indicate a good investment. The best investment would be a problem gambler.” Another former analyst was more blunt still: “It is as predatory as it sounds. If you lose more, we give you more, so you keep playing more.”

DraftKings disputes this characterization. The company told the Times it “rejects any implication” that its marketing unfairly targets customers, and says promotions are aimed at users showing “sustained, engaged” platform use. It also noted that Butts’s test analysis “appears preliminary and inconclusive” and that his views do not reflect DraftKings’ approach to promotional investment. The overlap between “engaged users” and “users predicted to lose the most,” however, is precisely the question — and it is one the company’s own modeling choices determine, since the model decides how heavily each feature is weighted.

The numbers executives bragged about

What makes the story harder for DraftKings to wave away is that the company itself has celebrated the commercial results of this data-science apparatus. Executives told the Times that data science and analytics improved promotion-driven sportsbook margins by roughly 13% in 2025, and that AI helped personalize hundreds of millions of dollars in promotional spending.

That is a striking number. A 13-point margin improvement attributable to promotional targeting is not a rounding error — it is one of the larger documented returns on applied machine learning in consumer tech. And it was achieved, according to the former employees, by optimizing promotional spend toward the customers whose predicted losses were largest.

The model that was shelved

The most damning thread in the investigation is not what DraftKings built, but what it chose not to finish. Former employees told the Times that a separate initiative — risk scores designed to identify customers showing signs of gambling problems — was stalled or shelved, even though the underlying data infrastructure to build it already existed. The pipelines could look in both directions; only one direction was ever productized.

DraftKings’ chief responsible gaming officer, Lori Kalani, told the Times that the company does monitor customers for risky behavior, and said DraftKings declined to deploy risk-prediction technology because it was “not sufficiently evidence-based.” The company publicly points to responsible-gaming commitments including a collaboration with Mindway AI’s Gamalyze tool, expanded customer education resources, and a partnership with IC360 for integrity and compliance monitoring. What remains unanswered is whether any of those tools can do what the shelved model was meant to do: proactively score and intervene with users at elevated risk, using the same behavioral signals that already power the promotional engine.

Why this matters beyond gambling

This story landed in the middle of the loudest AI-safety week in recent memory — frontier lab CEOs calling for slowdowns, Congress debating kill switches, and a fresh antitrust suit over coordinated pacing. Against that backdrop, the DraftKings investigation is a useful corrective. The most immediate AI harms are not hypothetical superintelligence; they are optimization systems quietly maximizing metrics that no regulator ever approved.

Every large consumer platform now holds the same raw material DraftKings held: dense behavioral telemetry, a predictive model, and an engagement objective. The pattern the Times documented — build the revenue-maximizing model, shelve the harm-detection model — is a governance failure, not a technical one. Nothing about the shelved risk-scoring model was harder to build than the loss-prediction model. It was simply less profitable.

For lawmakers and state gambling regulators, the investigation offers something more concrete than general industry unease: named sources, a specific 2023 model, quantified margin effects, and an on-record company denial that can now be tested. Several states have already been probing sportsbook promotional practices; a documented internal model that ranks customers by expected losses hands regulators both motive and material.

And for users, the practical takeaway is uncomfortable but simple. The free bet arriving in your inbox is not necessarily a reward for loyalty. It may be the output of a model that has quietly scored you — and decided you were worth the investment.

The uncomfortable truth inside the Times’ reporting is that the model worked. It found the losers, the losses grew, and the margins improved by double digits. The question every operator of a behavioral AI system should now expect to answer is not “what does your model do?” but “what did you decide not to build?”