When the Best Betting Ad Is No Ad at All

Key takeaways

• An advertising model’s goal shapes which users it learns to target.
• Optimizing for additional deposits can conflict with protecting bettors.
• Predicting betting after a loss does not establish that someone has an addiction.
• Responsible personalization needs rules for when to withhold an ad.

Imagine losing a bet, deciding to stop, and receiving a deposit offer just as your resolve starts to wobble. For a sportsbook such as DraftKings, that scenario raises a question more consequential than whether an ad matches your interests: what, exactly, is the system trying to get you to do?

Every optimization goal makes a choice

Personalized advertising changes the message depending on the person. Behavioral targeting uses information such as visits and clicks to help make that choice. AI can predict who is likely to respond.

But “respond” needs a definition.

A model optimized for clicks has a different assignment from one optimized for sign-ups or additional deposits. Change the success metric, and you can change which users the system seeks out.

Suppose the goal is a deposit after an ad. A model could learn to identify patterns associated with that outcome. The metric alone would not distinguish a planned deposit from an impulsive attempt to recover a loss.

Both can look equally successful on a marketing dashboard. The bettor may experience them very differently.

Predicting another bet is not diagnosing addiction

The idea that AI can “find people who will lose money” bundles together two separate claims: predicting who will lose, and predicting who will keep betting after losing.

The second task does not require predicting a game’s outcome. A system could pursue its advertising goal simply by estimating who is likely to deposit again after seeing an offer.

That distinction matters. So does the limit of what behavioral data can tell us.

Frequent visits or repeated deposits do not, by themselves, establish that someone has an addiction. A correlation is not a diagnosis. Yet uncertainty about a person’s psychological state does not settle whether using those signals to encourage more betting is acceptable.

Evaluating a sportsbook such as DraftKings would therefore require specifics: the model’s objective, the data it uses, and the timing of its ads. Saying a company “uses AI” tells us very little about whether its advertising responds to an interest or takes advantage of a vulnerable moment.

Timing is part of the offer

Consider a hypothetical bettor who has just lost ₩50,000 and immediately receives an offer encouraging another deposit.

Was the message part of a scheduled promotion, sent regardless of that loss? Or was it triggered because the system predicted that this person would be especially likely to deposit at that moment?

The wording could be identical. The decision behind it would deserve very different scrutiny.

An offer arriving after a loss could provide a reason to resume betting at precisely the point when someone might otherwise stop. Reading the ad copy alone would miss that possibility. The audience, timing, and intended response all belong in the assessment.

A useful question follows: could the same signal have prompted the service to give the bettor some space?

If the system could have withheld a promotion but instead used the signal to encourage another deposit, that choice reveals what its personalization serves.

Personalization can mean backing off

A personalized system can decide to send fewer messages—or none.

Possible design choices include pausing promotions during a break the user has chosen, or excluding users from campaigns when certain patterns of losses or repeated deposits appear. These are options for service design, not claims about DraftKings’ current practices.

The definition of success would also need to expand. More deposits alone cannot tell a company whether its advertising served the user well. Evaluation should also consider whether users stayed within their chosen limits and whether requests for a break were respected.

The boundary between useful marketing and exploitation depends on the choices built around a model. If a service can recognize a moment of vulnerability, it has a decision to make about how to use that knowledge. Sometimes the most responsible result is an ad that never gets sent.