A regulator-backed gambling risk model is now open source - but it is not a compliance shield

Online casinos already analyse deposits, stakes, session timing and losses. The unusual part of the Dutch project is that the resulting risk model is intended to be inspectable. The Kansspelautoriteit has published access to code and methodology developed by University of Amsterdam researchers, creating a reference outside a single operator's private system.

That changes who can ask meaningful questions. Researchers can examine assumptions, regulators can compare an operator's outputs with an independent baseline, and operators can test whether their controls miss patterns visible elsewhere. Open code does not guarantee good decisions, but it makes some decisions contestable.

The KSA says the algorithm considers how much and how often a person stakes, the frequency and timing of play - including repeated night-time sessions - and how behaviour changes after wins and losses. It converts those patterns into a risk score rather than waiting for one dramatic loss to trigger attention.

The University of Amsterdam says the training material comprised all wagers from all players at 13 Dutch online casinos between 30 July 2023 and 30 July 2025. That breadth is notable, but it should be described accurately: it is a large Dutch regulatory dataset, not proof of universal performance in every future market.

Behavioural models estimate patterns associated with risk. They do not diagnose a clinical disorder, establish why a person played at night or determine on their own what intervention is proportionate. False positives can burden players who are not in difficulty; false negatives can leave genuine harm unseen.

The meaningful questions are therefore about validation: how often the model is right, what kinds of risk it misses, whether performance changes across games and player groups, and what human review follows a flag. None of those questions disappears because the code is public.

The KSA explicitly says operators can use the model to improve duty-of-care work, but that doing so gives them no guarantees. It may only form an additional part of their measures. That sentence prevents a powerful but false compliance claim: 'we ran the official model, therefore our intervention was adequate.'

Duty of care includes what an operator does with information, not only how it scores information. Contact timing, affordability context, product controls, limits, escalation and record-keeping can remain relevant even when the initial risk detection is technically sophisticated.

Operators already possess detailed first-party data and often have proprietary models. A transparent reference gives supervisors a way to compare those systems without accepting their internal labels at face value. A large disagreement between scores can become the beginning of an audit question rather than an invisible modelling choice.

The UvA also says the project drew on collaboration with Spain's DGOJ. That cross-regulator lineage hints at a wider use: a common research baseline that can be tested across legal systems while local intervention rules remain distinct.

The code release is the opening, not the conclusion. Independent replication, clear error metrics, documentation of data limitations and evidence about real-world interventions will determine whether the model improves outcomes. Version history will matter too, because player behaviour and products change.

Atlas will treat future claims in layers. The model is public; its training scope is described; its use is optional and supplementary. Whether it becomes a supervisory benchmark, changes enforcement or reduces harm is not yet established.

Open source allows inspection, but most affected players will not read code. Operators and regulators still need plain-language explanations of what data is used, what a flag means, who reviews it and how a person can challenge a mistaken inference.

That is the differentiating story here. The Netherlands has not produced an automated answer to gambling harm. It has produced a public tool that can make hidden scoring systems easier to question - provided institutions publish the evidence needed to question how it is applied.