CSIPE

Published

- 19 min read

The Betting Offer Knows When You Are Likely to Say Yes

Privacy Artificial Intelligence Online Safety Advertising

Books by the author

Compare all 5

As an Amazon Associate I earn from qualifying purchases. Buying through these links costs you nothing extra and helps pay for the blog.

A free bet can arrive at exactly the wrong moment. Perhaps your team has just lost, your account balance is down, and you are still carrying the small, hot feeling that one more wager could repair the evening. The phone lights up with an offer shaped to make that next wager feel cheaper.

On 19 September 2026, The New York Times reported that DraftKings had used machine learning and customers’ betting records to predict who would lose more money after receiving a promotion. The report was based on internal material and interviews with former employees. DraftKings disputes the claim that it targeted people because they were likely to lose, and the public record does not yet settle every detail of how each customer was selected. (Operation Sports)

Five days later, the chair of the Massachusetts Gaming Commission directed staff to examine how licensed betting companies use artificial intelligence and machine learning, beginning with questions for DraftKings. That review is an inquiry, not a finding that the company broke a rule. It matters because a regulator has now turned a newspaper investigation into a concrete question for the industry: what should a betting company be allowed to predict about a customer, and what may it do with the answer? (CDC Gaming)

You do not need to understand machine learning to protect your next decision. You need to understand the gap between an offer that happens to appear and an offer selected because a system expects it to change your behaviour. Then you can put a little distance back where the app has tried to remove it.

What the September report actually says

The important allegation is narrower than the loudest version of the story. The Times reported that DraftKings built a model in 2023 that scored how a customer might respond to a promotion. A former analyst described the commercial question in plain language: would this person give the company more than the company gave them? According to the report, a favourable answer could lead to more promotional attention. (Operation Sports)

That does not mean a computer forced anyone to place a bet. It means the company could use a long record of earlier choices to estimate which offer was most likely to produce more activity and more losses. One person might ignore ten dollars in bonus credit. Another might deposit again, keep playing after a bad run, or return sooner than planned. A prediction becomes commercially useful when those differences can be measured across a large customer base.

The Electronic Frontier Foundation examined the report on 24 September and placed it in the wider category of behavioural advertising. This is advertising selected according to what a person has done, rather than the page they happen to be reading. EFF noted an especially important detail: the reported model could work with first-party data, meaning information DraftKings collected through its own customer relationship, without buying a separate dossier from an outside broker. (Electronic Frontier Foundation)

That distinction closes off an easy but incomplete response. Blocking third-party cookies or asking data brokers to delete a profile can reduce some tracking across the web. Those steps cannot erase the history held by a service you used directly. A betting operator already knows the deposits, wagers, wins, losses, withdrawals, sessions, bonuses, device events, and account controls recorded in its own system. The first-party record is enough to support detailed personalisation without an outside dossier.

The strongest claims still need careful wording. The Times investigation attributes the system’s purpose and use to documents and former employees. EFF’s article analyses that reporting; it does not provide a second set of internal records. The Massachusetts review confirms that the issue is serious enough for regulatory examination, while proving neither the newspaper’s full account nor DraftKings’ full response. As of 25 September 2026, anyone claiming that a regulator has already reached a verdict is getting ahead of the evidence.

What has been established is enough for a practical lesson. A company can use the detailed behaviour generated inside its own app to decide who gets a prompt and when. The recipient sees an invitation. The sender may see a calculated probability.

How a betting history becomes the next offer

Think of the model as a shopkeeper with a very long notebook. The notebook does not contain one grand secret about you. It contains many small observations: when you visit, which sports you follow, whether you respond to a bonus, how quickly you return after a loss, and what happened after the previous offer. A person could not study millions of such notebooks before sending tonight’s messages. Software can.

A machine-learning system begins with past examples. For each earlier promotion, the company can compare what was known before the offer with what happened afterwards. Did the customer open the message? Did they deposit? How much did they wager? Did their losses increase over the next day or week? The system looks for combinations that helped predict the measured outcome in that historical data.

The result may be a score rather than a sentence. A customer might receive an estimate of how much additional activity or revenue a particular offer could produce. Marketing software can then rank customers, select an offer, or decide which people should receive nothing. The message still looks friendly and personal because the machinery stays behind the screen.

Here is how that can play out in an ordinary evening. Sam plans to spend twenty dollars on a match and stop. The first bets lose, so Sam deposits another twenty. A previous promotion led Sam to continue playing for an extra hour, and the account history contains similar weekends. If a model has learned that pattern, a timely bonus may be sent because Sam is likely to respond, not because every football fan received the same deal.

The prediction does not have to know why Sam responds. It may combine hundreds of signals whose relationships are hard to describe in one clean rule. Perhaps time of day matters. Perhaps a recent deposit, the type of bet, the speed of play, or a pattern of previous bonus use matters. Some signals will be meaningful; others may be accidental stand-ins for circumstances the company never named.

This is one reason the word “AI” can distract from the useful question. Whether the scoring system is called artificial intelligence, machine learning, data science, or customer analytics changes little for Sam. The decision that matters is whether the company uses a prediction of likely harm or likely loss to reduce pressure, or to increase it.

There is also a feedback loop. If the system sends more offers to people who respond, their records produce more evidence about response. Future models are then trained on a history partly created by earlier targeting. The system can become very good at finding the behaviour it has repeatedly encouraged, even if nobody wrote a rule saying “contact this person after a difficult night.”

A prediction can be wrong, of course. Models confuse coincidence with cause, rely on incomplete records, and change as customers change. False predictions do not make the practice harmless. A person wrongly classified as highly responsive may receive unwanted pressure. A person correctly classified may receive a prompt precisely when resistance is low. Either way, the model’s uncertainty is carried by the customer, while the company can spread its bets across millions of decisions.

Why “personalised” means something different here

Personalisation can be genuinely useful. A music app remembers what you enjoy. A supermarket reminds you about an item you buy every month. A bank may recognise an unusual payment and ask whether it is yours. Treating every tailored service as equally harmful would leave us with a weak argument and worse products.

Betting promotions have a different alignment of interests. A bookmaker’s revenue depends in part on customers losing money over time. A customer may enjoy the entertainment and accept a planned cost, but the operator and customer do not want the same result from each wager. When a promotion is chosen according to who is expected to spend or lose more, the friendly language covers a sharper commercial calculation.

The timing matters as much as the content. An offer for a winter coat does not normally arrive during a state in which buying coats becomes hard to control. Gambling can. A message sent after a loss, during a long session, or when someone has repeatedly returned may meet a person whose decision is already being pulled by frustration and the hope of recovery. The same ten-dollar offer can be trivial at lunchtime and powerful at midnight.

People sometimes answer this by saying adults remain responsible for their choices. They do. Personal responsibility and commercial responsibility can exist in the same room. We still regulate how credit is advertised, how medicines are promoted, and how companies treat sensitive information because a market becomes less fair when one side can study a person’s weak moments at industrial scale.

The private asymmetry is the heart of the issue. You can see the bonus amount and the expiry time. You usually cannot see the score that selected you, the outcome being optimised, the other offers tested on similar customers, or the behaviour that moved you into this audience. The company gets a detailed view of the pattern. You get a notification that feels like a lucky break.

That asymmetry also makes consent difficult. A privacy notice may explain that data supports personalisation and marketing, but few people can translate that sentence into a picture of a loss-prediction model. Clicking “accept” at account creation does not give a customer an honest sense of the later decision. The gap grows when a new use is developed years after the original data was collected.

DraftKings also offers responsible-gaming controls, including limits and cool-off periods. On 14 September 2026, the company announced custom breaks ranging from three to 364 days, alongside new responsible-engagement material. Those controls are useful and should be used when needed. Their existence does not answer how promotional models should behave around the same customer. A safety brake deserves protection from the marketing engine, especially when the engine can recognise the conditions that make braking difficult. (DraftKings announcement carried by Yahoo Finance)

The privacy setting that helps, and the one it cannot fix

Use the privacy choices offered in your account or regional privacy form if they are available to you. Also review phone permissions, browser tracking settings, and advertising choices. These steps can reduce the number of companies passing identifiers and browsing behaviour around the advertising market.

The limit is written into the shape of the problem. A cross-site advertising opt-out concerns data used across different businesses or contexts. The September reporting focused on betting records created inside DraftKings itself. A company does not need a tracking cookie from a news site to know what happened in its own sportsbook.

Turning off phone notifications has a similar boundary. It stops the app from placing a message on the lock screen. It may not stop an offer appearing inside the app, arriving by email, or influencing which bonus appears after login. Unsubscribing from promotional email can close another door, but account or legally required messages may still arrive.

These are reasons to use several small controls rather than search for one perfect privacy switch. Remove push notifications. Unsubscribe from marketing email and text messages. Exercise the privacy choices offered for targeted advertising. Then set a money or time boundary inside the betting account. Each control blocks a different route.

The order matters. A limit chosen while you are calm is stronger than a promise made while an offer is counting down. Decide the amount before the match, set it in the account, and avoid changing it during the same session. If you find yourself repeatedly raising the limit, the useful signal is the repeated change, not the number you finally chose.

Browser tools deserve a modest place in the plan. Tracker blockers can reduce third-party surveillance around the web, and phone settings can reduce advertising identifiers. They cannot turn a service with an account history into a stranger. Do not spend an afternoon tuning obscure cookie lists while leaving the app free to contact you at the moment you most want distance.

This is the same sensible middle used throughout The Digital Fortress: Your Everyday Guide to a Safer Digital Life. Spend your patience where it buys you the most safety. Here, the useful work happens in the account controls, the phone’s notification settings, and the payment boundary you can check later.

A promotion is designed to change the decision

“Free” betting offers often feel like money found on the pavement. Their real purpose is to change what happens next. The company can afford the offer when the expected return from changed behaviour exceeds its cost. A promotion that never changes a deposit, wager, or return visit would be a poor marketing expense.

This does not make every promotion deceptive. A clearly explained bonus may provide exactly the value it promises. The safer question is personal: would I be placing this bet, at this amount and at this time, if the offer had not appeared? If the answer is no, the promotion has already done its job.

Countdowns add pressure by turning a commercial choice into an apparent deadline. The event may begin soon. The bonus may expire. A boosted price may have a clock beside it. None of those clocks knows whether the wager fits your budget, and none deserves authority over a decision that can wait.

Try a cooling rule that the app cannot shorten. Any unexpected offer sits untouched until the next day. For an event happening tonight, that means missing the offer. This is a feature of the rule. A boundary that preserves every promotional opportunity has been designed around the promoter’s needs.

The rule also produces useful evidence. After a night’s sleep, you can ask whether the offer still looks valuable or whether it mostly looked urgent. You may notice that the excitement belonged to the timing, not the terms. That observation is more useful than arguing with yourself while the notification is still glowing.

Family members can help without becoming police. If someone wants support, agree on a simple check before deposits above a chosen amount, or ask the bank whether gambling blocks are available. Keep the agreement specific and voluntary. Shame drives decisions underground; a calm second pair of eyes creates the pause that targeted messages try to remove.

A person who is worried about losing control needs more than better ad settings. In the United States, the National Council on Problem Gambling provides call, text, and chat routes through 1-800-MY-RESET, and local services vary by state. Use the current contact details on its site rather than relying on an old number saved in an article. Help is appropriate before a crisis, not only after one. (National Council on Problem Gambling)

What to actually do before the next notification

Start while there is no match on, no balance to recover, and no offer expiring. The goal is a short setup you can verify, rather than a heroic promise to make perfect decisions under pressure. Ten quiet minutes can change the next difficult evening.

  1. Turn off the invitation routes. Disable betting-app notifications in the phone’s settings, then unsubscribe from promotional email and text messages. Open the app once more to check its own communication preferences. Save a screenshot of the settings, because an account migration or reinstall can change them later.

  2. Choose a money boundary in advance. Set a deposit limit that fits money already allocated for entertainment. A weekly or monthly limit is usually easier to understand than a per-deposit maximum, which may still permit several deposits. Use the longest period that makes your total spending obvious.

  3. Add a time boundary. Decide when the session ends, regardless of whether the account is up or down. Use the app’s time controls where available and set an ordinary phone alarm outside the app. The outside alarm matters because it remains visible even when the betting interface is trying to hold attention.

  4. Make unexpected offers wait. Give any unplanned promotion a one-day cooling period. Do not deposit to use it, and do not replace it with a similar offer elsewhere. If the event passes, let it pass. Missing a bonus costs less than letting someone else’s prediction choose the evening.

  5. Use a stronger barrier when the small ones keep moving. A cool-off period temporarily blocks access. Self-exclusion is stronger and may apply across operators within a jurisdiction. A bank gambling block can add another independent layer. Choose the barrier according to what has actually been happening, not according to how serious the label sounds.

  6. Check the result, not the intention. One week later, confirm that promotions stopped, limits remain in place, and no second account or payment route has replaced the first. A screenshot, confirmation email, or bank setting is a receipt. “I meant to cut back” is a hope; a boundary you can inspect is evidence.

If you do only two things, silence the promotions and set the deposit limit. You have my blessing to ignore the maze of advertising-industry preference pages until those two are done. They put distance between the prediction and the money, which is the distance that matters tonight.

Someone supporting a partner or family member should resist taking over every account unless invited or required for immediate safety. Begin with one observed pattern and one agreed control. “The offers seem to arrive after you have lost” opens a conversation. “You cannot be trusted with your phone” closes it.

Keep the practical goal modest. A person does not have to prove that a particular message was selected by a machine. They do not have to identify the exact score or win a debate about artificial intelligence. They can decide that betting promotions do not belong on the lock screen and make that true.

What companies and regulators should ask next

The Massachusetts review creates an opportunity to ask for facts rather than promises. A regulator can examine which outcomes promotional models optimise, what customer data enters them, how often they are tested, and whether markers associated with harmful play are excluded from targeting. It can also ask whether the responsible-gaming team has the authority to stop a campaign, rather than merely advise the marketing team.

The first hard question is about purpose. “Personalisation” is too broad. A model designed to show baseball offers to baseball fans differs from one designed to find customers whose losses rise after a bonus. Both are personalisation, but they carry different risks and deserve different controls.

The second question is about separation. If one system estimates who is likely to respond profitably and another estimates who may be experiencing harm, what happens when the same person scores highly in both? A sensible design gives the safety signal priority and prevents the promotional system from reaching that customer. The exclusion should be enforced in software and tested, rather than left as a sentence in a policy.

The third question is about proof. Companies should be able to show which model and policy selected a promotion, what version was running, which data fields were used, and whether a safety exclusion was checked. That record should survive long enough for a complaint or audit. A customer-facing explanation can stay simple while the regulator receives enough detail to test the claim.

Independent evaluation matters because the company benefits from a model that increases activity. An internal dashboard may celebrate deposits, engagement, and revenue while failing to count debt, distress, or the number of people who repeatedly override limits. Regulators should choose measures that reflect the public purpose of gambling rules, not accept the campaign’s success metric as the safety metric.

The public should also know the limits of the current evidence. As of 25 September 2026, the Massachusetts commission has announced examination, not a completed enforcement case. The Times has published serious reporting, and DraftKings contests the harmful interpretation of its work. A careful review should test the disputed facts and publish enough reasoning for customers to understand the result.

One policy principle does not need to wait. A company that can predict vulnerability should not use that prediction to apply more pressure. If the model cannot reliably distinguish ordinary interest from a harmful pattern, that uncertainty argues for less targeted promotion, not more.

Put the pause outside the app

The September story is about DraftKings, but the mechanism belongs to a much larger family. Shopping apps predict who will buy after a discount. Video services predict which image will earn a click. Delivery apps learn which reminder brings someone back. Most of those nudges cost attention or a purchase you may regret. Betting adds the possibility that the system learns from losses and reaches a person during an attempt to recover them.

The calm response is not to assume every offer is a trap. Treat every personalised offer as somebody else’s proposed plan for your next action. You are allowed to decline the plan without proving why it reached you.

Privacy controls help, especially when they reduce sharing across companies. Notification settings help because they protect the quiet moment before you open an app. Money limits, cooling periods, self-exclusion, and bank blocks go further because they constrain the transaction itself. Use the lightest control that keeps its shape, then move to a stronger one when you find yourself bending it.

The best boundary sits outside the moment it must govern. Set the limit on a quiet morning. Turn off the message before the match. Ask for help before the account reaches the number that scares you. A model may be able to predict when you are likely to say yes, but it cannot press a button your environment has already removed.

For one calm, practical security briefing each month, join the newsletter. One email per month, and no noise between them.

Sources