Dutch Regulator Funds Open-Source Algorithm to Score Player Risk

Researchers at the University of Amsterdam have built a machine-learning model that reads betting behaviour rather than what players say about themselves, and released it free. The Kansspelautoriteit funded it and is careful to call it an instrument rather than a substitute for duty of care.

• Researchers at the University of Amsterdam have released an open-source machine-learning algorithm that produces a risk score for individual online gambling accounts
• The work was backed by the Kansspelautoriteit's Addiction Prevention Fund, the Dutch regulator's own harm-prevention funding
• The model reads activity data: betting behaviour, gambling frequency, session timing, and winning and losing streaks alongside how players react to them
• The KSA's stated position is that the algorithm is another instrument, not a substitute for human oversight or the broader duty of care
• The team drew on earlier work by Spain's gambling regulator, the Directorate General for the Regulation of Gambling, when developing the approach

A Regulator-Funded Model That Anyone Can Download

A machine-learning model designed to identify signs of risky online gambling has been released as open source by researchers at the University of Amsterdam, funded through the Kansspelautoriteit's Addiction Prevention Fund.

The team is PhD candidate Charles de Leau, working with Professor Reinout Wiers of the psychology faculty and Professor Johan Bollen of computer science. Their approach drew on earlier work by Spain's gambling regulator, the Directorate General for the Regulation of Gambling, known as the DGOJ.

The model works from what players do rather than what they report. It examines betting behaviour, gambling frequency and the timing of sessions, and looks at winning and losing streaks alongside how a player responds to them, producing a risk score for each account.

That design choice is the substantive one. Self-report instruments depend on a player recognising and disclosing a problem, and risk frequently emerges through changes in behaviour that a player does not recognise, or would not disclose if asked. Activity data does not have that dependency.

Releasing it openly changes what it is. Commercial player-protection tools are proprietary, which means an operator buying one cannot fully inspect how it reaches a conclusion, and a regulator reviewing that operator's decisions inherits the same opacity. A published model can be examined, criticised and tested by anyone, including the people it scores.

The KSA Is Careful About What It Is Not

The regulator's framing is deliberately narrow. The KSA's position is that the algorithm is another instrument, not a substitute for human oversight or the broader duty of care, and it is intended for use by operators as an additional layer within their existing player-protection systems rather than as a replacement for them.

That caution is doing real work, because Dutch duty-of-care obligations sit on the operator. A tool funded by the regulator and offered to the market could easily be read as a compliance standard, and the KSA has been explicit that it is not offering one.

A Free, Inspectable Model Puts Commercial Vendors on the Wrong Side of an Argument

The responsible-gambling technology market runs on proprietary scoring, sold on the strength of a vendor's data and validation rather than on published method. An openly available model backed by a national regulator does not have to outperform those products to change the conversation. It only has to be good enough to make "why is your system better than the free one that the regulator paid for" a question every procurement process now asks. Vendors with genuine performance advantages can answer it with evidence. Vendors selling opacity as sophistication have a harder time, and the pressure will be sharpest on the ones whose pitch has always leaned on the difficulty of building this in-house.

The Real Question Is What Happens When an Operator Ignores a Score

The KSA says this is an instrument rather than a duty-of-care substitute, and that is legally the right thing to say. It does not settle how the tool will be used in a dispute. Once a free, regulator-funded model exists that would have flagged a given account, an operator that either did not run it or ran it and did nothing is answering a different question than before: not whether its monitoring was reasonable in the abstract, but why an available and endorsed signal did not produce an intervention. That is a heavier evidential position than the current one, and it arrives without any rule having changed. Operators would be wise to decide now, and document, how they will treat scores they choose not to act on.

Behaviour-Based Scoring Cuts Across the Affordability Debate

Most regulatory attention on player protection has gone to affordability, which asks what a customer can lose, and answers it with financial data that operators find intrusive to gather and customers resent providing. This model asks a different question, about how someone is playing rather than what they can afford, using data every operator already holds. If behavioural scoring proves reliable, it offers regulators a route to harm detection that does not require bank statements, which is politically easier and commercially far less damaging to conversion. That makes this more interesting to jurisdictions currently fighting the affordability argument, Britain included, than a Dutch research release would normally be.

The KSA has funded a tool and declined to make it a rule. Whether the market treats it as one is now largely out of the regulator's hands.