The recent release of the open-source risk model developed by the University of Amsterdam (UvA) and endorsed and funded by the Dutch Gaming Authority (KSA) has reignited an essential conversation in player protection. Making machine learning models and two years of online multi-casino consumer data accessible to regulators, researchers, and operators is an important step toward transparency in safer gambling.
However, as the industry works to establish reliable, standardised frameworks for player protection, we must examine a foundational question: What are these models actually measuring?
There is a critical distinction between predicting an action such as a player hitting a self-exclusion button and detecting underlying harm or shifts in human behaviour. Relying on self-exclusion as a main proxy for gambling disorder introduces significant measurement errors that limit a system’s ability to protect players early and effectively.
The Scientific Dilemma: A Measurement Problem, Not a Modelling One
The distinction between an administrative tool and a diagnostic marker is equally clear from a neuroscientific perspective. Self-exclusion was designed as an immediate circuit breaker for players, requiring no formal assessment or diagnostic screening to activate. Using an unverified, self-selected event as the proxy for a clinical condition introduces significant noise into training datasets.
Kim Mouridsen, Founder, Mindway AI & Professor of Neuroscience, Aarhus University:
“Self-exclusion was built as a safety handle and not a diagnostic marker. It is a convenient way for anyone to step away quickly, without any assessment required. Those are the right properties for a protective tool but the wrong ones for an early detection system. Using self-exclusion as a surrogate marker for harm is therefore scientifically controversial at best and this is demonstrated in the literature. Studies show that among online gamblers, most who self-exclude do not have a gambling disorder. More problematically, around 80% of those who do and need help never self-exclude at all.
This is a measurement problem, not a modelling one. The use of machine learning to predict self-exclusion is not new and has been studied for around 10 years. The problem remains that high technical accuracy in self-exclusion prediction does not equal early detection of people in need of help, and scientifically it tells us little about the early course of harm-and on that the literature is far larger, strikingly convergent and already embedded in clinical practice.
Therefore, at Mindway AI we inverted the problem. Rather than learn from a self-selected event, we have clinical psychologists score real, anonymised player behaviour across a broad set of markers and train the AI to reproduce that expert reasoning. Because any player can be scored, our training data is large and randomly drawn-not confined to the minority who press a button.”
The Clinical Limitation: Why Self-Exclusion Misses the Target
If raw data alone is insufficient, using self-exclusion as the target variable for machine learning creates a deeper issue. In clinical psychology, self-exclusion is understood as an endpoint or a crisis coping mechanism, not a reliable marker of progressive harm.
Relying on self-exclusion to train AI models creates a fundamental paradox: the system learns to recognise players who eventually exclude themselves but remains blind to those who suffer severe harm without ever taking that step.
Annika Lindberg, Chartered Psychologist & Mindway AI Expert Panel Member:
“The UvA/KSA model is essentially very good at answering one particular question: what patterns of gambling behaviour are associated with someone going on to self-exclude? Self-exclusion is undoubtedly correlated with gambling problems and many of the behavioural features used are highly relevant to risk. The question is what we can infer when someone is flagged within this framework.
Is self-exclusion a good enough marker for what we are actually trying to identify; namely emerging gambling harm?
If self-exclusion becomes the main marker against which risk is learned, we may become very good at recognising people who eventually self-exclude, while being less sensitive to those whose gambling is becoming increasingly harmful but who continue to gamble. Clinically, I see many people with severe gambling problems who never self-exclude. Poor insight, feeling conflicted about stopping and remaining caught in the chase to recover losses can all keep people gambling despite significant harm. Conversely, self-exclusion does not necessarily indicate that someone is experiencing harm from gambling.
This matters particularly for early detection. Ideally, we want to identify meaningful behavioural change before someone reaches an obvious endpoint. Self-exclusion is a useful signal, but it should not stand in for the much broader construct of gambling harm. A more rounded approach would therefore look for multiple, clinically informed changes in gambling behaviour, rather than relying on any single outcome as a proxy for harm.”
A Lived Experience Perspective: Looking at the Gaps in Data vs. Human Context
From a practical and operational perspective, algorithmic tracking of betting frequency, stakes and loss patterns is not inherently new. Most modern gambling operators utilise data models to track player activity. What sets the open-source model apart is its accessibility, not its core mechanics.
Yet, relying strictly on raw transaction data leaves crucial blind spots regarding human context, environment, and neurodiversity amongst other things.
Mark Potter, MP Consulting Ltd:
“It has been interesting to read the new, free to access UvA/KSA open-source risk model, based on consumer data from two years of play from thirteen casinos. Undoubtedly, there is a place in the safer gambling/player protection space for us to understand more about player behaviour from machine learning. In fact, most online operators already use algorithms that analyse betting patterns, time and frequency of bets placed and win/loss patterns, so in many ways, this is nothing new or groundbreaking. It is just, to my knowledge, the first time it has been made free to access for researchers, regulators and the general public.
Looking at it from the lens of lived experience and having spent many years both as a disordered gambler, and for the last 13+ years working extensively across the safer gambling training, education and consultancy space, my overriding feeling, is that there are some key variables that are missing or yet to be addressed with the risk model in question.
Factors such as circumstances, environment, personal composition, neuro diversity and emotional wellbeing all make up a significant part of how someone consumes gambling, which then undoubtedly feeds into any behavioural and motivational change. Using a personal example, as an athlete, you are judged by success and failure and when I gambled and lost, I found this difficult to accept and the answer for me was always to practice harder and get better. Then through periods of injury, I used gambling as a method of filling the void from lack of competition, which changed my motivation to gamble from fun and enjoyment to filling a void. I was also diagnosed with ADHD later in life post treatment, which again played a part in my relationship with gambling and ability to make good choices in times of stress.
In summary, a model based on only data will always have some gaps that will need to be expanded on by truly understanding individual behaviours, circumstances and motivations.”
The Broader Impact: Standardising Metrics That Work
As European regulatory bodies move toward unified standards such as the CEN (European Committee for Standardization) Markers of Harm project the accuracy and definition of underlying risk models become matters of policy and public health. When evaluating national or international standards, defining “risk” by a self-selected action risks instilling a false sense of security within regulatory frameworks.
Rasmus Kjaergaard, CEO, Mindway AI:
“When we today comment on this open-source initiative in the Netherlands we do it based on all the experience we have gained since working with AI models for monitoring of gambling behaviour trained by gambling behaviour psychologists. This is experience gained from investing thousands and thousands of working hours on training models, building software and not least deploying them with gambling operators all over the world. A combined pool of experience from operating in 73 license jurisdictions actively tracking over 16.5 million monthly active players’ individual gambling behaviour across varying operator sizes, game verticals and regulatory frameworks worldwide. Experience gained as a leading provider in the Dutch market of monitoring solutions since the Dutch online market opened 5 years ago and now working with the biggest Dutch gambling operators and counting.
Picking up on the very important points made by Mark about behavioural context, it is important that any model is able to review multiple risk factors to be able to provide some insights into what is happening with the player. GameScanner has 15 separate models designed to do just this. The “practice and get better” element would show up as loss chasing, the increased activity to fill the void would show up on the time consumption risk factor and the erratic decision making would be covered by impulsive transaction behaviour. All of these would have impacted the risk score much earlier and enabled interaction and later on, intervention before problems escalated to the point of self-exclusion.
Further, it’s experience gained by being a leading monitoring solution provider being actively involved in the CEN (European Committee for Standardization) Markers of Harm project, helping shape standardised, evidence-based protection metrics across Europe. The outcome of this EU standardising prestige project heavily relies on the highest possible accuracy and efficacy of the underlying model and if that’s not available, one could consider if it even makes sense to standardise Markers of Harm. It may likely be a misconception if one claims that a model trained to predict self-exclusion increases player protection, our experience shows for the reason outlined above rather the opposite.”
Conclusion: Moving Beyond the Single-Proxy Paradigm
Relying on self-exclusion as a standalone proxy for gambling harm introduces a fundamental measurement flaw into predictive modelling. Because self-exclusion is a self-selected administrative action rather than a clinical status, using it as a primary target metric yields high margins of error flagging low-risk players who simply wish to take a break while completely missing the remaining individuals suffering severe, unaddressed harm who never hit the self-exclusion button.
Building an effective, proactive player protection framework requires moving away from single-event proxies. True early detection depends on a multi-dimensional approach that evaluates nuanced behavioural trajectories, incorporates human clinical expertise and accounts for the lived realities of players. Open-source initiatives offer a valuable foundation for collaboration, but the industry’s ultimate goal must extend beyond predicting when a player will reach a breaking point, it must focus on identifying subtle behavioural changes early enough to prevent that harm from occurring in the first place.
Advancing safer gambling technology is an evolving, industry-wide conversation. If you have perspectives, research, or ideas on refining behavioural assessment models and moving beyond single-proxy metrics, Mindway AI CEO Rasmus Kjaergaard welcomes the dialogue. We invite researchers, operators, and regulators to connect and explore how we can collectively elevate player protection standards.
Get in touch contact@mindway.ai
