France's risk algorithm flagged 600,000 accounts linked to 60% of tracked gambling losses

France's gambling regulator says an algorithm applied to account-based play identified around 600,000 people with a high probability of excessive gambling in the second half of 2025. They represented 8.7% of the player population covered by the model.

The same group generated €1.2 billion, equivalent to 60% of the gross-gaming-revenue base described by the ANJ. In the regulator's July clarification, that figure is framed through player losses over the semester, with the small number of positive-balance players removed from the calculation. It is a concentration measure, not total stakes.

The ANJ explicitly says the tool is not designed to count the exact number of people with excessive-gambling problems or replace population-prevalence surveys. It assigns a risk score from behavioural and financial indicators observed in account data.

Those indicators cover money movements, use of gambling limits, activity, frequency and player history. The resulting categories run from recreational to moderate risk, excessive and manifestly excessive. A high-probability classification can trigger scrutiny and intervention without becoming a clinical finding about a named person.

Operators reported progress: the number of excessive players they identified rose from 31,000 in 2024 to 89,000 in 2025. The regulator nevertheless says that total remains inconsistent with the size of the player base and the evidence from prevalence work.

About 300,000 of the modelled players were placed in the most severe category. The ANJ expects operators to identify that manifestly excessive group in the short term and to improve detection across the wider population of roughly 600,000.

The model is available to operators on an optional basis as a complement to their own systems. The ANJ will use it as a point of reference for monitoring trends and assessing whether detection efforts match observable risk in account data.

Detection is only the first control. The regulator lists calls, tailored limits, referrals to support or care services and account closure among possible responses. A larger flagged population will matter only if interventions are timely, proportionate and evaluated for effect.

The 2027 review of operator prevention plans could make the comparison concrete by placing operator totals beside the regulator's benchmark. Until then, the sharpest responsible headline is the gap already documented: a model found probable excessive play at a scale many times greater than the cases operators reported identifying, and that group accounted for most of the measured loss base.

Data governance will be part of that test. A model built from continuous account information can improve consistency, but false positives, missed cases and changes in player behaviour still need supervision. Operators need a route for trained staff to interpret alerts, record decisions and escalate risk without treating an automated score as the final word about a person.

The concentration number also creates a commercial question that cannot be ignored. If harm-reduction measures work, revenue from the flagged group should fall. A market that treats that decline as a control success, rather than simply a lost sales line, will be better aligned with the regulator's stated objective.