DraftKings ranked casino players by how much they would lose after a free bet

A model learns what its label tells it to learn. Here the label was dollars lost after a promotion: the training set was bettors who had received an offer, followed by a measurement of how much they subsequently lost ^[8]. Tech Times, summarising the investigation, argues that the inputs DraftKings chose are the same behavioral markers a clinical AI gambling-prediction study uses to identify disordered gambling. A model fitted to maximise predicted post-promotion loss will therefore rank users by how closely they resemble those markers, it says ^[9].

The people who built it did not need the argument. "We are looking for traits and features that we can target that indicate a good investment," Jayden Butts, a former DraftKings data analyst who tested the system on thousands of casino players, told the Times ^[5]. "The best investment would be a problem gambler," he said ^[6]. A second former analyst, unnamed, told the paper that "It is as predatory as it sounds" ^[7].

The counter-thesis is that expected post-promotion loss also ranks a wealthy recreational player who loses steadily and does not much mind, because the label is a dollar amount and not a diagnosis ^[8]. The Times did not report any measured overlap between high elasticity scores and self-exclusion, helpline contact or clinical assessment, so the diagnostic version of the claim rests on the description given by the analysts who built the model ^[4].

Nestor Hernandez's crisis-detection project, begun in 2024, would have run on the same data pipeline that fed the promotional system, according to people familiar with the work ^[12][13]. That makes the incremental cost mostly headcount. Lori Kalani, the chief responsible gaming officer, told the Times that leaders had reached a "collective decision" against predictive risk tools. The approach was deemed not "evidence-based" ^[14]. DraftKings instead integrated Mindway AI's Gamalyze, a simulated card game the customer has to sit down and play, into its Responsible Gaming Center as an alternative diagnostic ^[15]. The shelved model would have scored behavioral data the company already held.

In my view the marketing-spend exposure is the smaller half of this. Promotional targeting is a marketing practice, and marketing practices get regulated forward from the rule's effective date; the shelved project and the "evidence-based" rationale are documented internal knowledge, dated 2024, and that is what a plaintiff's expert works with ^[12][14].

Scale, for context. The National Council on Problem Gambling's 31,000-plus monthly helpline contacts work out to about 372,000 a year ^[16][1]. Set that against the roughly $400 million DraftKings ran through automated promotions in 2025 and you get about $1,075 per annual helpline contact ^[10][3]. The two populations are not the same people, and the helpline covers every form of gambling. Treat it as a size check. On the demand side, the 2026 American Sport Fanship Survey put loss-chasing among online sports bettors at 60%, up from 52% ^[17]. That is an 8-point move and a 15% relative rise in one year ^[2]. And 26% said losses are causing them financial problems ^[18].

If the top elasticity decile turns out to be mostly high-income players who lose without distress, the clinical framing weakens to ordinary price discrimination. If the 13% margin credit came from withholding promotions from unprofitable accounts, the AI story is cost control ^[11]. And if state regulators treat loss-sensitivity targeting as ordinary marketing, the 2025 promotional budget stands as priced.