DraftKings AI Model Targeted High-Loss Gamblers, Problem Gambling Warning Project Shelved

On September 19, 2026, The New York Times reported, based on interviews with more than 40 former employees, internal memos, and betting records, that DraftKings has trained machine learning models since 2023 to score users by a “resilience score,” using it to direct promotional bonus bets precisely to users expected to lose the most.

Factual Reconstruction

The model’s training objective was to predict how much users would subsequently lose after receiving free bets or bonuses, with input features including gambling frequency, daily account balance patterns, and the ratio of losses to total wagers. These features closely overlap with behavioral markers used clinically to identify problem gambling. In 2025, the company distributed about $400 million in AI-automated promotions through the system, and management said it increased promotion-driven sports betting profit margins by 13%. A problem gambling early-warning model initiated in 2024 by data scientist Nestor Hernandez was shelved because the company deemed it to “lack an evidence base,” and the company instead adopted the Mindway AI Gamalyze tool.

Mechanism Breakdown

The resilience model’s training data label was “actual amount lost after receiving a promotion,” so the model naturally prioritized identifying users who continued to bet frequently and accumulate losses even under the stimulus of promotions. Former data analyst Jayden Butts said the system was looking precisely for the traits of “problem gamblers.” The same data pipeline could have supported a risk-warning model, but due to company decisions it was used only to maximize promotional returns.

Industry Impact

This case shows that AI compliance issues lie not only in the model itself, but also in how companies choose to use it. DraftKings says its promotions target “consistently active users” and that it has integrated more than 24 behavioral indicators for responsible gambling interventions, but former employees point out that shelving the warning model made the internal data infrastructure serve only profit maximization.

Strategic Assessment

[Analysis] When companies possess the technical ability both to identify vulnerable users and to identify at-risk users, how decision-makers define an “evidence base” will directly determine where AI resources flow. While similar cases previously focused mostly on model bias, this case provides a complete chain of evidence showing that usage decisions themselves can amplify harm, offering a reference point for future compliance frameworks in gambling and technology regulation.