Researchers at the University of Amsterdam have released an open-source machine-learning model designed to identify risky behavior among online gamblers, giving regulators an independent way to compare the player-protection systems used by licensed operators. The model was commissioned by the Netherlands Gambling Authority, or Ksa, and became publicly available on Aug. 18.
The algorithm analyzes actual gambling behavior rather than relying primarily on surveys or information supplied by players. Its inputs include how much and how frequently someone bets, the times at which they gamble, repeated nighttime sessions, winning and losing streaks, and how behavior changes following those streaks. It then uses those patterns to calculate a risk score for the player.
University of Amsterdam PhD candidate Charles de Leau developed the model with psychology professor Reinout Wiers and computer science professor Johan Bollen. The researchers trained it using all wagers placed at 13 licensed Dutch online casinos between July 30, 2023 and July 30, 2025, taking advantage of a provision in Dutch gambling law that allows operator data to be made available for independent research.
The two-year dataset gives regulators an alternative to risk models developed internally by gambling companies or commercial suppliers. Operators already collect detailed information about deposits, wagers, session length and other customer behavior, but the systems used to decide when that activity becomes dangerous are generally proprietary. The Ksa can now calculate its own risk scores and compare them with the assessments produced by operators.
Dutch licensed operators have a legal duty of care requiring them to monitor gambling behavior and intervene when signs of excessive play or addiction appear. The Ksa has repeatedly found problems with how that responsibility is handled, including a €750,000 fine issued to ComeOn operator Tulipa Ent in January after the regulator found failures in all 10 high-loss young-adult player files it examined. Bet365 operator Hillside was separately ordered to strengthen its response to signs that customers could no longer afford their gambling.
The Ksa has made operator monitoring one of its supervisory priorities for 2026 and said earlier this year that it planned additional guidance around AI and automated monitoring. Chairman Michel Groothuizen said the regulator would examine not only losses but also which behavioral indicators operators monitor and how quickly they intervene after detecting them.
The new model does not automatically determine whether an operator has met its legal obligations. The Ksa says licensed companies can use the algorithm themselves, but deploying it is only one possible part of a wider player-protection system and does not guarantee compliance with the duty of care. Human intervention, affordability checks and other safeguards remain separate requirements.
The researchers also worked closely with Spain’s Directorate General for the Regulation of Gambling, which has been developing its own approach to identifying risky players. Because the University of Amsterdam code and methodology are public, regulators outside the Netherlands can inspect the model, test it against local data and modify it rather than buying a closed commercial system.
Other regulators are already applying AI and large-scale data analysis to different parts of gambling oversight. Colombia’s Coljuegos launched an AI center in 2024 that monitors illegal gambling websites, identifies cloned versions of licensed operators, tracks social media promotion and watches transactions from authorized gambling machines. That Colombian system is separate from the Dutch player-risk algorithm, but both reflect regulators moving toward automated analysis instead of relying entirely on operator reports and manual investigations.
The University of Amsterdam model now gives regulators a common reference point for one of the harder parts of online gambling supervision: deciding when ordinary play starts showing signs of escalating risk. Operators will still make their own intervention decisions, but the Ksa no longer has to judge those systems solely through models designed and controlled by the companies it regulates.