DraftKings is using AI to behaviorally target chronic gamblers
TL;DR Highlight
DraftKings trains an ML model on betting records to target gamblers likely to lose with ads.
Who Should Read
Backend and data developers building targeting or recommendation models who worry about ethics and regulatory risk.
Core Mechanics
- DraftKings picks out likely losers
- Per the New York Times, it trains a machine learning model on customers' betting records.
- Users the model flags get targeted ads designed to lure them back to bet more.
- The losing users are the ones who pay
- EFF says DraftKings has a business incentive to keep losing gamblers coming back.
- It argues problem gamblers are highly likely to be targeted by this model.
- AI amplifies behavioral ad harms
- Training and refining models requires collecting even more data.
- EFF says AI is a black box, so builders can rarely predict which data points matter and keep collecting more.
- Ad data feeds the surveillance industry
- EFF says ad-targeting data is sold to insurers, banks and law enforcement such as CBP.
- Earlier this year ICE published an RFI on how ad tech providers could directly support investigations.
- First-party data is enough
Community Reactions
- A commenter shared a ProPublica piece on DraftKings' response to problem signs
- The reporter worked with gambling addiction experts to test how the company responds.
- The commenter concluded the company has no ethics.
- A commenter's grad course covered casino churn models
- Caesars Entertainment modeled its most valuable customers and predicted when they were about to leave.
- It intervened with perks, and competitors rushed to copy it.
- A commenter cited a Connecticut study on revenue concentration
- The 2024 study found 1.8% of problem gamblers accounted for 51% of sports betting revenue.
- The commenter was surprised the target is chronic losers, not whales.
- Commenters called for regulation
- One wants ID checks on every login and device-level self-ban, like casino floors.
- Another described an ecosystem of affiliates, data brokers, credit bureaus and CRMs profiling potential addicts before signup.
How to Apply
- Audit who your targeting model targets
- Today, pull the target conditions of your churn-prediction or re-engagement models in CRM, push or recommendations.
- Check whether users showing loss, overspending or addiction signs fall into the target set.
- Add an exclusion rule for vulnerable users
- In sensitive domains like gambling, payments or subscriptions, add exclusion conditions at the targeting step.
- Log the number of excluded users and the reason.
- List the purpose of each collected field
- List the first-party fields used for model training.
- Mark each as needed for the service or only for ad optimization.
Try It
# Paste into Claude or ChatGPT as is (a simplified checklist prompt based on the article's argument)
Review our targeting/retention model from an ethics and privacy perspective.
[Service description]
{{service description}}
[Model purpose]
{{e.g., churn prediction then push/coupon sending}}
[Data fields used]
{{field list}}
Summarize in a table:
1. Paths by which the model could target vulnerable users (overspending, addiction signs, etc.)
2. Risks that exist even with first-party data only
3. Suggested exclusion conditions
4. Fields we could stop collecting
Next: ask it to turn the exclusion conditions into actual query or filter code.Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.