Massachusetts regulators find sportsbooks unable to explain their own systems

Massachusetts AI probe reveals sportsbooks cannot explain their own systems

Dean Serpa sent a five-page letter to seven licensed mobile operators on October 9, and the question buried in section six is the one that matters: if a regulator asked why a specific patron received a specific promotion on a specific day, could the operator identify the data and model behind that decision? The November 3 deadline will answer that, and I think the answer most operators return is going to be no.

The Massachusetts Gaming Commission's review was triggered by a New York Times report alleging DraftKings used machine learning to target customers most likely to lose. DraftKings denied wrongdoing. The MGC's response was to look at every licensed operator, not just DraftKings, which is the correct instinct and also the more revealing one. A probe that starts with one company's alleged conduct and then asks all seven licensees the same questions is really asking whether the problem is one bad actor or one bad industry practice.

Serpa's letter defines AI broadly enough to catch anything an operator might try to relabel — "personalization," "optimization," "risk assessment" all fall inside it. That definition alone tells you the commission anticipated pushback on terminology. What the letter then asks is whether losing bettors are treated differently from winning ones in the distribution of promotions, and it follows that with a harder question: if the answer is no, what controls exist to prevent it? That sequencing is deliberate. It is harder to claim you don't do something if you also cannot describe the mechanism that would stop you.

Commissioner Paul Brodeur noted last month that Massachusetts is one of the only states to reference AI directly in regulation. That regulatory head start is now an investigative advantage. The MGC has a legal basis to demand answers that most other state bodies don't. Michigan's Gaming Control Board said it is reviewing the Times report, but Michigan's laws don't address AI, which means whatever Michigan finds, it has limited tools to act on.

The reporting describes participation as voluntary. That framing should not be mistaken for toothlessness. A licensed operator that declines to respond to a regulatory inquiry in a state where it holds a licence is making a decision with consequences that are easy to anticipate. Voluntary in letter, not in practice.

Here is where I break from the obvious read: most coverage of this story frames it as a consumer protection issue, a question about whether vulnerable gamblers are being targeted. That is accurate but secondary. The deeper problem the MGC has uncovered is operational opacity. If operators genuinely cannot answer why Patron A received Promotion X, the issue is not just ethics — it is that these systems are making material business decisions that the businesses themselves cannot audit. A regulator sitting across from an operator that cannot explain its own model is not looking at a compliance gap. It is looking at a control failure, and those tend to have consequences that extend well beyond the original investigation.

The human oversight section of the letter presses on exactly this point. At what stage does a person review the model's recommendation? Can that person override it? Do they receive an explanation of why the model flagged a particular patron? Those questions assume the answer might be no, and they are structured to make a no-answer visible rather than defensible.