A gambling model can learn to reward vulnerability without ever having a field called “addiction.” That’s the uncomfortable engineering lesson in the reporting on DraftKings—not some sci-fi story about a rogue chatbot.
I get excited about predictive systems. Turning messy behavioral data into useful decisions is genuinely fascinating. But when “useful” means finding customers expected to lose more money after a nudge, the dashboard needs more than a cheerful conversion chart.
It needs brakes.
The scoreboard nobody should celebrate
A September 19, 2026 New York Times investigation by Alex Klavens, Walt Bogdanich, and Jenny Vrentas reported that DraftKings began developing a promotional machine-learning model in 2023. Drawing on internal documents and former employees, the investigation described customer scores reflecting expected losses after receiving incentives.
Those incentives include free bets, profit boosts, and deposit bonuses. Familiar marketing machinery. A disturbing optimization target.
Former data analyst Jayden Butts, who tested the model, questioned whether its most financially attractive customers were also vulnerable to addiction:
<> Under strict financial logic, the “best investment would be a problem gambler.”/>
That’s his assessment of the incentives—not proof that every highly ranked customer has a gambling disorder. The distinction matters. Predicting promotional profitability is not the same as clinically identifying addiction. The Hacker News headline compresses that distinction too aggressively.
The reported stakes are substantial: Citizens Bank research cited by the Times estimated $8.7 billion in gross sports-and-casino revenue and approximately $3 billion in promotions during 2025. Those are gross figures, not financial-statement revenue after deductions.
The Real Story
The interesting question isn’t “Did they use AI?” It’s which objective got engineering resources, and which objective got veto power?
Six former employees involved in promotional targeting described continued model refinements. Four other former employees alleged that DraftKings stalled or suppressed efforts to predict gambling problems using similar technology. These are reported allegations, not adjudicated findings.
If that imbalance is substantiated, it’s the heart of the story. A company can build both a growth model and a safety tool while allowing only one to influence the money-making decisions.
On September 14—five days before the investigation—DraftKings announced responsible-engagement initiatives featuring Kevin Hart and Nick Jonas, Gamalyze American Football with Mindway AI, and cooling-off periods of three to 364 days. Chief Responsible Gaming Officer Lori Kalani described responsible engagement as embedded in operations.
Useful context. Not evidence that promotional models reliably exclude vulnerable customers, and not a direct rebuttal to the subsequent investigation.
A celebrity campaign cannot substitute for an eligibility check.
First-party data, first-class risk
[EFF’s analysis](https://www.eff.org/deeplinks/2026/09/draftkings-using-ai-supercharge-harms-online-behavioral-advertising) highlights something developers should notice: this targeting apparently uses first-party betting records.
No data broker required. No cross-site tracking circus.
An authenticated product already knows what its customers do. That means restrictions on selling or sharing data would not necessarily stop this particular mechanism. EFF argues for banning behavioral advertising altogether; that’s a broader policy proposal, not the only remedy established by the evidence.
My position: high-risk promotional systems need enforceable safety constraints, not optional safety features.
Put the veto in the serving path
We don’t know DraftKings’ exact architecture, features, or deployment rules. Nothing here establishes an LLM. These are engineering recommendations, not claims about its implementation:
1. Name the objective honestly. “Engagement” should not conceal expected customer losses.
2. Make safety exclusions binding. Cooling-off status and validated risk flags should constrain promotional eligibility across channels.
3. Audit harmful outcomes. Ranking accuracy alone cannot tell you whether a system disproportionately reaches vulnerable customers.
Keep decision records, too:
- Model versions and feature provenance.
- Campaign assignments and eligibility rules.
- Human overrides and who authorized them.
That’s less glamorous than shipping a clever model. It’s also where accountability becomes testable.
I still love what machine learning can do. This story is a reminder to inspect what we ask it to do. If the success metric rewards customer harm, better predictions make the problem worse.
