DraftKings Used AI to Find the Gamblers Most Likely to Lose, Then Targeted Them

The clearest documented case of AI being used to exploit rather than serve consumers arrived on September 19. Per The New York Times’ investigation, DraftKings used machine learning models built by its data science teams to identify customers most likely to lose money gambling, then targeted those losing gamblers with free bets and promotions encouraging more betting. A predictive tool meant to detect problematic gambling was reportedly shelved or squashed during development.

The ML Application Is Standard; the Objective Is Not

What makes this investigation notable for ML practitioners is that the underlying technique is unremarkable. Customer lifetime value prediction, propensity modeling, and churn prevention are standard data science across every consumer business. DraftKings simply inverted the objective: instead of identifying customers to retain, the models identified customers to exploit. The same feature engineering, the same gradient boosting, the same A/B testing infrastructure. The ethics live entirely in the objective function, not in the architecture, which is why “responsible AI” frameworks that focus on model fairness and explainability miss this category of harm entirely.

The Shelved Detection Tool Is the Damning Detail

The investigation reports that DraftKings developed a predictive tool to detect problematic gambling and then shelved it. That decision converts a story about aggressive marketing into a story about deliberate harm enablement: the company built the capability to identify at-risk users and chose not to use it, while simultaneously deploying models to find and target them. In regulatory terms, this is the difference between negligence and intent, and it is exactly the pattern that state attorneys general and the FTC look for when deciding whether to pursue consumer-protection actions.

The Broader Pattern for AI Practitioners

This investigation, the DraftKings CEO touting AI optimization in February, and the regulatory vacuum around AI-driven consumer exploitation together illustrate where AI harm is actually heading: not rogue agents or existential risk, but optimization toward extraction in the absence of a duty of care. The Frontier Act’s focus on frontier-model oversight does not touch this, the G20 principles do not touch this, and the California AI safeguards signed this month address chatbots and deepfakes but not propensity modeling for exploitation. The gap between what AI can do and what the law prohibits it from doing is widest not at the frontier but in the commercial applications that are already deployed, already profitable, and already invisible to the people they harm.