https://doi.org/10.1007/s10899-026-10483-6
COMMENTARY

The Need for Benchmarks to Advance AI-Enabled Player Risk
Detection in Gambling
Kasra Ghaharian1 · Simo Dragicevic1 · Chris Percy1 · Sarah E. Nelson2 ·
W. Spencer Murch3 · Robert M. Heirene4 · Kahlil Simeon-Rose1,5 · Tracy Schrans6
Received: 17 December 2025 / Accepted: 5 February 2026
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2026

Abstract
Artificial intelligence-based systems for player risk detection have become central to
harm prevention efforts in the gambling industry. However, growing concerns around
transparency and effectiveness have highlighted the absence of standardized methods for
evaluating the quality and impact of these tools. This makes it impossible to gauge true
progress; even as new systems are developed, their comparative effectiveness remains
unknown. We argue the critical next innovation is developing a framework to measure
these systems. This paper proposes a conceptual benchmarking framework to support the
systematic evaluation of player risk detection systems. Benchmarking, in this context,
refers to the structured and repeatable assessment of artificial intelligence models using
standardized datasets, clearly defined tasks, and agreed-upon performance metrics. The
goal is to enable objective, comparable, and longitudinal evaluation of player risk detection systems. We present a domain-specific framework for benchmarking that addresses
the unique challenges of player risk detection in gambling and supports key stakeholders,
including researchers, operators, vendors, and regulators. By enhancing transparency and
improving system effectiveness, this framework aims to advance innovation and promote
responsible artificial intelligence adoption in gambling harm prevention.
Keywords Responsible gambling · Machine learning · Artificial intelligence ·
Benchmark datasets · Problem gambling · Prevention

Kasra Ghaharian
kasra.ghaharian@unlv.edu

International Gaming Institute, University of Nevada, Las Vegas, NV, USA

Division on Addiction, Cambridge Health Alliance, Harvard Medical School, Malden, MA,
USA

Department of Psychology, University of Calgary, Calgary, Canada

Brain & Mind Centre, School of Psychology, Science Faculty, University of Sydney, Sydney,
Australia

Carson College of Business, Washington State University, Everett, USA

Focal Research Consultants, Halifax, Canada

Introduction
Gambling is a significant and growing form of entertainment, participated in by millions
globally. While most consumers enjoy gambling as a leisure activity, prevalence rates for
harm vary considerably across different populations, jurisdictions, and product types. Globally, an estimated 1.4% of adults meet criteria for problem gambling, but these rates shift
dramatically depending on the gambling environment; for example, prevalence rises to
approximately 16% among online casino or slots gamblers (Tran et al., 2024). The industry’s rapid expansion, driven largely by digitization and the global availability of online
and mobile products, has led to a corresponding increase in efforts to mitigate these potential harms (Ghaharian et al., 2022). These collective efforts are commonly referred to as
Responsible Gambling (RG)1.
RG strategies have evolved as a direct result of gambling’s digitization, and include
targeted RG messaging, immediate access to gambling help resources, and automated risk
detection systems (Shaffer et al., 2020). These technological player protection measures can
be divided roughly into two types: user-initiated measures and automated operator-initiated
strategies. The former includes tools that users can employ, usually voluntarily, to control
their gambling: self-limiting features, reality checks, deposit limits, cool-off periods, and
self-exclusion. However, few users take advantage of these features (Ladouceur et al., 2016;
Nelson et al., 2008), and there is debate about whether those who do are the players who
need them most (Catania & Griffiths, 2021).
While automated, operator-initiated strategies can also take many forms, including universal approaches like deposit limits, there has been a movement in the past decade toward
more tailored interventions (Heirene & Wohl, 2025). Universal strategies risk either disrupting the activity of those who do not need protections or, if implemented in a way that does
not disrupt most users’ activity, being set so conservatively that many of those who need
them do not trigger them (Broda et al., 2008). On the other hand, targeted strategies can, in
theory, use player information to better determine when and to whom to deploy protections.
As such, data-driven RG models have become the holy grail of automated operatorinitiated RG solutions (Ghaharian et al., 2022). These systems often leverage artificial intelligence (AI) – more specifically, machine learning (ML) – to analyze players’ gambling
behavior and detect patterns associated with harm to trigger interventions. This approach
has attracted substantial interest from both academia and industry stakeholders, with several
regulators mandating the use of data-driven systems for player risk detection, operators
developing proprietary models, and specialist technology vendors emerging to offer commercial solutions.
However, a fundamental question remains largely unanswered: Are these sophisticated
systems effective?
This paper argues that the next critical innovation required is not the advancement of current methods or the creation of new models, but the development of a framework to measure
them. We propose that this problem can be addressed by applying Performance Benchmarking, a common practice in many AI domains, but one that is yet to be applied in this use case.

We use the term Responsible Gambling (RG) throughout this paper as it is commonly used to describe these
practices. We acknowledge the ongoing and important discourse contrasting the RG framework to safer
gambling or public health approaches, but a full discussion of that debate is beyond the scope of this paper.

The paper is presented as follows: We first provide a brief background on the use of ML
for player risk detection, followed by a discussion of the current challenges in evaluating
these systems. We then introduce the concept of Performance Benchmarking and briefly
review its application in other domains. Finally, we present a conceptual, discussion-starting
framework for a player risk detection benchmarking suite. Rather than proposing a single,
fully specified benchmark, we outline the structural requirements, core dimensions, and
practical considerations necessary for such benchmarks to be developed and adopted by the
research and practitioner communities. Our hope is to stimulate discussion and thought for
what such a framework would require and what it could achieve.

Machine Learning for Player Risk Detection
A ML approach to player risk detection typically uses information about gambling patterns
and player characteristics from gambling records (behavioral markers) to forecast potential
gambling-related problems2. This is similar to the use of biomarkers in medicine, where
objectively measured characteristics serve as indicators of disease processes (Biomarkers
Definitions Working Group, 2001). Medicine also uses behavioral markers, for example, in
early detection of Alzheimer’s disease (Gomar et al., 2011).
ML models can target players according to markers defined by both their gambling patterns and changes in those patterns, using this information to designate differing levels of
potential risk. Most importantly, these models can provide a guide for matching responses
to individuals, optimizing the opportunity for favorable outcomes. Very basic best practices
for developing such models include establishing that the markers differ in the groups to be
distinguished (e.g., individuals who exhibit problematic gambling behavior and those who
gamble without problems), developing a predictive model that integrates markers to predict
the outcome in question, and then establishing the model’s performance in samples separate
from those used to develop the model (Morrow et al., 2008; Vanderstichele et al., 2008).
One distinct feature of ML is its ability to learn complex, nonlinear relationships in data,
making it highly suited for identifying at-risk play given the individual variability in gambling behavior (Hastie et al., 2009). However, complex ML solutions encounter some of the
same challenges as the most basic models, such as establishing accurate outcome measures
and training data and developing models that can generalize and replicate beyond the idiosyncrasies of a specific dataset (i.e., distinguishing the signal from the noise: Silver, 2012;
also see Shaffer et al., 2020).
Early player risk models re-purposed data from loyalty smart cards used in physical
gambling venues (e.g., casinos) (Schellinck & Schrans, 2011). The shift to online gambling,
however, significantly simplified behavioral tracking. Unlike physical venues that depended
on a player inserting a card, online platforms and mobile apps inherently capture comprehensive player data. This seamless data collection fueled a new research area, leading to
at least 40 academic studies and several literature reviews (we refer the interested reader
to: Ghaharian et al., 2022; Delfabbro et al., 2023; Marionneau et al., 2025). Alongside this
academic work, commercial models were developed. A recent review by Marionneau et
al. (2025) identified 11 commercial models, including six in-house systems from operators
(e.g., Kindred, Veikkaus) and five from third-party vendors (e.g., Mindway, Openbet).

For an up-to-date and comprehensive review of model architectures and the current landscape, we refer the
reader to Marionneau et al. (2025)

These models vary widely in their methods, employing a variety of unsupervised and
supervised algorithms, and in the number and type of input markers they use. As Marionneau
et al. (2025) note, these markers are often unclear or inconsistently reported in the literature,
and the number of markers used in a given model range from fewer than 10 to over 100.
Ghaharian et al. (2025b) recently mapped the markers in the academic literature, identifying
65 “low-level” categories (such as bet amount and net loss) which can be grouped into five
“high-level” categories: play (betting behavior), payment (financial transactions, i.e., deposits
and withdrawals), engagement (when, how often, and how broadly players engage with gambling), RG tool use, and profile information (e.g., registration date, demographic information).
While literature reviews have helped map the academic literature, comparing model performance remains very difficult. Studies often use different datasets, model parameters, and
data cleaning procedures, making it hard to draw clear conclusions about which methods
are most effective. This challenge is compounded by the significant transparency problem
with commercial models, as the information available on these systems is insufficient for a
meaningful comparison. These and other challenges are discussed in the following section.

Current Problems
It is appropriate to hold ML to high standards, given that policy increasingly references
it in supporting harm reduction (Bedford & Chandan, 2024). Moreover, such standards
align with the sensible and realistic performance beliefs that gamblers themselves may hold
(Murch et al., 2025). Several challenges limit the potential for ML in this context.

Broad Challenges for Stakeholders
ML, as with all identification and intervention design, involves trade-offs between sensitivity and precision, i.e., where to place the balance between failing to identify gamblers
at risk and falsely identifying many gamblers with negligible risk levels (Murch et al.,
2023). These trade-offs can be improved with more sophisticated models but cannot be
fully resolved in real-world settings. In the absence of precision, an ultra-low specificity
model simply provides RG interventions to all players, which may undermine trust in RG
programs and potentially lead to substitution by players towards other operators or products.
As models grow more complex, they typically become less transparent with respect to
how any individual decision has been reached (Sarkar et al., 2016). This is often referred to
as the “black box” problem, where the model’s internal logic is so intricate that it becomes
opaque even to its creators, making it difficult to audit or explain a specific outcome (von
Eschenbach, 2021). Explainability, or the ability to identify how a model arrives at its prediction, is important to facilitate users’ trust in a system and ensure the model is operating as
expected (Sarkar et al., 2016). In addition, explainability has practical relevance: it enables
operators to activate appropriate interventions and allows players to reflect on behavioral
dimensions that have been identified as increasing their risk.
Another issue is that models can only identify patterns in the data they possess. Consequently, identical gambling behaviors can represent very different risk-levels depending
on variables not included in a data set, such as an individual player’s life circumstances or
financial context. For example, the potential impact of a specific monetary loss is difficult

to contextualize without information about an individual’s financial circumstances (e.g.,
income, debt exposure, etc.). Moreover, harms are not purely financial. A significant amount
of time spent gambling could be a harmless hobby for one person, but for someone suffering
psychological distress, it could be a coping mechanism leading to other harms like relationship disruption or missing work. Much of this relevant contextual information simply
cannot be collected by operators under current social norms. This missing context is why
seemingly identical profiles based on available data could have different risks and, as the
UK Gambling Commission (2023) has noted, why high-risk players exist at all stake levels.
There are also challenges related to AI ethics generally, such as risks of bias in datasets
or models (Percy et al., 2020), data collection protocols, and ensuring human oversight for
high-consequence decisions. While a full discussion of AI ethics in gambling is beyond
this paper’s scope (discussions available in Binesh and Ghaharian, 2025; Ghaharian et al.,
2025a; Percy et al., 2022), we note the trend towards translating ethical principles into regulation (e.g., the EU AI Act approved in May 2024).

Specific Challenges for Procuring or Commissioning Models
Those procuring or commissioning models for player risk detection face practical questions:
which model performs best by a given metric (e.g., sensitivity, precision, F1 score), which
metric is most important to consider, how each model performs across different player
groups, and what interventions are attached to the model’s classifications. These questions
are valuable for both procurement managers and those commissioning in-house tools.
Given the quantitative nature of data science, it might seem surprising that these questions are not easily addressed today. This void is often filled by competing corporate marketing claims. For example, claims about risk detection model accuracy were discussed at
the International Association of Gaming Regulators annual conference (Percy, 2024), with
examples including claims that models can exceed 80% accuracy (Sustainable Interaction,
n.d.), 87% (Mindway, n.d.), and over 90% (Entain, 2022).
While including such numbers is a positive step for corporate transparency, these claims
are nearly impossible to evaluate without critical context for at least three reasons. First, a
“high accuracy” claim can be misleading. For example, in a population where only 10% of
individuals have a problem, a model could achieve 90% accuracy simply by identifying no
one as at-risk. Second, these figures are not necessarily comparable. It is not possible to conclude that a 90%+ model is definitively better than an 87% model, without context to ensure
we are comparing like with like. Key factors such as the definition of risk, the prevalence
of risk in the underlying sample, and the approach to out-of-sample testing (i.e., validating
the model on new, unseen data) can all lead to dramatically different performance results,
even when using the same modeling technique. Finally, these claims are not independently
reviewed, and it remains unclear whether other models’ performance remains unreported for
reasons of privacy or poor performance.
Reviewing published academic papers or patent filings provides additional data beyond
marketing and website claims but does not provide the full detail necessary (Marionneau et
al., 2025). While additional metrics and methodological detail are available, it is rare to find
exactly the same reporting criteria across studies, simply because there are so many metrics
and methodological choices involved. More seriously, even where full detail is available
across contexts, there are no methods to adjust resulting performance metrics for true like-

for-like comparison across those different contexts. Generalizability in player risk profiles
and model structures is limited across brands (Ghaharian et al., 2024), across time (Murch et
al., 2024b), and likely across a host of other factors such as gambling type, jurisdiction and
regulatory context. As a result, it is insufficient to lean solely on better norms for consistent
reporting3. More transparent and standardized reporting is only one step towards answering
a reasonable, fundamental question from end users: How does one model’s performance
truly compare to others?

System Consequences
For operators, the inability to effectively compare, evaluate, or audit player risk detection
systems creates a paradox. They can claim adequate RG measures without revealing specifics of their systems. However, this makes it difficult for good actors to demonstrate to regulators and the public that these measures are truly high quality. In the absence of industry
standards or independent benchmarks, some operators may base decisions on cost-effectiveness and ease of integration rather than accuracy and efficacy, leading to market preference
for low-cost, easily deployable solutions above model quality. This approach can cause
reputational damage and incur regulatory action, especially if a major incident occurs4.
There are no perfect comparative metrics, and we should recognize the limits in so-called
“league table” approaches, which attempt to create an oversimplified ranking of models
from best to worst. However, this does not mean comparison attempts should be shelved.
The gambling sector might reasonably look with envy at comparative metrics used in
other AI-driven fields. For instance, established benchmarks for tasks like translation, image
labeling, and Large Language Model (LLM) performance (e.g., intelligence, coding, agentic capabilities) are widely used5.
To be clear, these existing benchmarks are not perfect and are themselves the subject of
ongoing research to improve their design (Bean et al., 2025; Hardy et al., 2024). But they
prove that useful, non-simplistic comparisons are possible. Such comparisons support sector
decisions and may fuel model improvements across the board, as they have done in some AI
sectors, while building confidence and trust with stakeholders. At present no such benchmarks
exist for player risk detection systems, so the only way of knowing which risk model works
best in a given context is to build, buy, and apply them all—a profoundly impractical option.

Benchmark Datasets and Performance Benchmarking
A fundamental practice in ML research is evaluating algorithms against standardized
benchmark datasets (Longjohn et al., 2024; Thiyagalingam et al., 2022). Across numerous
domains, benchmark datasets have proven successful in identifying optimal ML solutions

For examples of standardized reporting guidelines and checklists for ML models and AI-based systems, see
TRIPOD-AI (Collins et al., 2024), CONSORT-AI (Liu et al., 2020), SPIRIT-AI (Cruz Rivera et al., 2020).

For example, the UK Gambling Commission fined Corbett Bookmakers £0.7m in March 2025, where one
identified issue was “failing to identify a consumer who staked £23,674 in a 13-day period as someone who
may be at risk of or experiencing harms associated with gambling” (​h​t​t​p​s​:​​​/​​/​w​w​​w​.​g​a​m​b​l​i​n​g​​c​o​m​m​i​​s​s​i​o​​n​​.​g​​o​v​
.​​​u​k​/​n​​​e​w​s​/​​a​r​t​i​​c​​l​e​/​g​b​​​p​6​8​6​​​-​0​7​0​​​-​f​i​n​​e​​-​f​o​r​-​c​​​o​r​b​​e​t​t​-​b​o​o​k​​m​a​k​e​r​s​-​l​i​m​i​t​e​d).

See, for example, public aggregations of these benchmarks at epoch.ai/benchmarks and artificialanalysis.ai.

for specific problems by enabling cross-stakeholder collaboration and fostering competitive environments for innovation. This process, Performance Benchmarking, allows for
objective quantitative comparisons between different approaches, as methods and results
cannot be reliably compared when models are trained on different data and evaluated with
different diagnostics (Thiyagalingam et al., 2022). Such standardization helps track progress over time, assess novel methods, and select the most appropriate approach for a given
application. Importantly, benchmark datasets should serve as proxies for real-world tasks,
so performance on the benchmark is generalizable to practical applications (Longjohn et
al., 2024). Beyond scientific validation, benchmarking can also serve as a critical component of regulatory infrastructure, providing a transparent and accountable framework for
evaluating complex systems (Percy et al., 2022). To understand how this framework could
be applied in the gambling field, it is instructive to review its successful implementation
in other domains.

Background
One of the earliest and most well-known examples of a benchmark dataset is the Modified
National Institute of Standards and Technology (MNIST) database (LeCun et al., 2010).
Created in 1998, MNIST consists of 70,000 size-normalized, 28x28 pixel images of handwritten digits. Each image is accompanied by a label indicating the correct number. Before
MNIST, tracking progress on numeric recognition was difficult because researchers used
different and incompatible datasets. MNIST provided a large, standardized collection that
allowed researchers to test and benchmark their models against a consistent baseline.
Throughout the 2000s, numerous benchmark datasets emerged, applying MNIST’s
fundamental principles to a growing range of use cases and more complex applications.
For example, the Street View House Number (SVHN) dataset, created in 2011, contains
600,000 labeled digits cropped from Street View images (Netzer et al., 2011), and sought
to advance image recognition beyond simple handwritten digits by providing a platform
for identifying characters in natural environments. Similarly, the MedMNIST benchmark
was created to advance biomedical image classification in healthcare (Yang et al., 2023).
More recently, the rapid advancements of LLMs have given rise to numerous benchmarks to assess their capabilities. Many, like the Massive Multitask Language Understanding (MMLU) test6 consist of large sets of multiple-choice questions across various domains
(Hendrycks et al., 2020). Similar to how SVHN advanced image recognition, newer LLM
benchmarks move beyond multiple-choice questions to assess more practical skills. For
example, Vending-Bench uses a simulated environment to evaluate LLM-agents’ ability to
operate a vending machine business (Backlund & Petersson, 2025)7.

MMLU is designed to measure knowledge across diverse domains (e.g., humanities, social sciences, hard
sciences) by evaluating a model’s accuracy on multiple-choice questions; for details, see its dataset card at
https://huggingface.co/datasets/cais/mmlu.,

Vending-Bench is a simulated environment that tests how well an AI agent can manage a long-running business scenario. The agent must make decisions about inventory, orders, and pricing in order to make money
(https://andonlabs.com/evals/vending-bench).

Types of Benchmarking
Benchmark datasets are as varied as ML itself and can be categorized along several axes. At
a fundamental level, they can be distinguished by data type: images, audio, text, time-series,
and tabular data. They can also be characterized by the ML task they address: classification,
detection, regression, etc. The data curation method can also characterize benchmarks, with
datasets ranging from purely synthetic data (e.g., see ACT-Thor: Hanna et al., 2022) to realworld data (e.g., SVHN).
Beyond these characteristics, benchmarking can be categorized as scientific-based or
competition-based. As described by Dueben et al. (2022), scientific benchmarks provide a
platform for addressing broad, long-term problems shared by many research groups, with
performance often self-reported through peer-reviewed publications. In contrast, competition benchmarks generally focus on narrower tasks that benefit from short-term, intensive
focus of a wider community. Kaggle is a well-known example where monetary awards are
often given for the top-performing solution. For instance, the “American Express – Default
Prediction” competition challenged participants to use anonymized transactional data to
predict which customers would fail to pay their debts (Howard et al., 2022). Other platforms, such as HuggingFace, Papers with Code, the UCI ML Repository, and OpenML, also
facilitate benchmarking (typically without monetary components) by providing infrastructure to host data, share models, and track performance on public leaderboards. As such, the
line between these categories can blur.

The State of Benchmarking in Player Risk Detection
Despite the success of benchmarking in other fields, its application in gambling studies
remains nascent, with no formal benchmark datasets for player risk detection. The closest
example is research stemming from The Transparency Project8 which made several datasets
publicly available. The bwin dataset, in particular, functioned as a de facto benchmark for
researchers interested in player risk detection. The bwin dataset consists of player account
records from a European-based online sports betting provider. A subset of these accounts is
labeled with the reason for account closure, including “gambling problems,” which commonly serves as a proxy for problematic play. This dataset enabled foundational ML studies, from initial cluster analyses (Adami et al., 2013; Braverman & Shaffer, 2012) to later
classification tasks comparing various supervised ML models (Philander, 2013). However,
the dataset’s utility is limited by its age and the fact that it represents data from only a
single operator. Although ML studies have continued to emerge since The Transparency
Project, these efforts remain fragmented. Research groups often produce valuable work, but
it frequently relies on proprietary data that cannot be publicly shared for broader validation
(Ghaharian et al., 2022).
Thus, a primary barrier to establishing benchmarks in gambling studies is data accessibility. Because player data is commercially sensitive and proprietary, it creates perceived
challenges in data sharing and research transparency. Although early initiatives like The
Transparency Project partially overcame these issues, their impact was limited by the
technological constraints of the time. More recently, the University of Nevada, Las Vegas

For details on this initiative, see The Transparency Project’s website: http://www.thetransparencyproject.
org/,

(UNLV) Payments Research Collaborative adopted a more modern data-sharing model,
publishing a peer-reviewed paper on a dataset of gambling payment transactions and making analytic subsets available on a public repository (Ghaharian et al., 2023). Similarly,
Zendle and Newall (2024) contributed by releasing a synthetic dataset9 derived from the
banking transaction data of 424 UK-based survey respondents, linked with their Problem
Gambling Severity Index (PGSI) scores. Still, without operator cooperation, these efforts
are limited in their generalizability.
Regulator-coordinated requests have proven effective for targeted research: in the UK,
the Gambling Commission worked with GambleAware to commission player account data
across seven major operators, although barriers limited the availability of the data to third
party researchers (Forrest & Mchale, 2022). Additionally, legislative mandates could offer a
more structural approach. For example, the Massachusetts Expanded Gaming Act requires
casino licensees to provide anonymized customer tracking data to the Massachusetts Gaming Commission, which must make this data available to qualified researchers (Andrews et
al., 2025).
Beyond data accessibility, creating benchmark datasets for player risk detection presents
several other complex challenges. For example, a benchmarking solution must tackle the
definition of the target outcome variable, as the very concept of harm in gambling can be
ambiguous, contentious, and is inherently multifaceted. Furthermore, it must be capable
of integrating diverse data sources, from wagering activity to payment transactions, and
account for the distinct behavioral patterns across different gambling verticals, such as
sports betting, online casinos, and land-based venues. In the following Section “A Proposed
Benchmarking Framework for Player Risk Detection Systems”, we attempt to initiate a
conceptual discussion by outlining the foundational components that a robust benchmarking framework would likely require, before examining its associated challenges in Section
“Challenges and Considerations”.

A Proposed Benchmarking Framework for Player Risk Detection
Systems
A comprehensive benchmarking solution will require a broad suite of labeled datasets representing dimensions necessary to model risky player behaviors. The solution should provide
sufficient datasets to enable industry participants to train and test risk detection models
across these dimensions—we refer to this as the benchmarking suite. For this paper we
describe three core dimensions: Time, Engagement Level, and Gambling Vertical.

Core Dimensions
A conceptual model of the benchmarking suite outlines some core dimensions that could be
captured across multiple datasets (Fig. 1).
Time refers to the length of the period of play that the data spans. For explanatory purposes in this paper, we sub-categorize time into sessions, days, and months. Providing a

A “synthetic” dataset is an artificially generated dataset. In this case, the authors used their original dataset
to generate a new dataset that mimics the statistical patterns and correlations of the original. This process
allows researchers to share data for analysis without exposing any of the original, privacy-sensitive data.

Fig. 1 Core dimensions of the benchmarking suite

short-term view (i.e., sessions) is crucial for identifying “in-the-moment” risk behaviors,
such as within-session loss-chasing (Parke et al., 2015). Longer-term horizons, such as
days or months, are necessary to understand cumulative impacts and identify at-risk players within specific time-frames. For example, prior studies have sought to assess future risk
using as little as 7 days (Auer & Griffiths, 2024) or 30 days of data (Braverman & Shaffer,
2012), while others have used 12-month datasets to create ML models (Kairouz et al., 2023;
Murch et al., 2023).
Engagement Level may refer to the frequency, regularity, and/or historical duration of
a customer’s play. This dimension is critical as it directly impacts the nature of the data
available for modeling. For instance, if we consider a dataset with a longer time horizon
(e.g., 12-months) the behavioral patterns and available data for a long-term, regular player
will be vastly different from those of a new or infrequent player, the latter representing a
data-sparse problem with only a few sessions. From a ML perspective, it is likely that different models, features, or thresholds are required to effectively detect risk in these distinct
cohorts. Additionally, ML models deployed without regard to engagement levels might do
little more than distinguish those who play regularly from infrequent gamblers, instead of
identifying those patterns that are specifically associated with gambling problems or harms
among regular gamblers. Thus, a robust benchmarking suite would need to contain datasets
that represent the range of engagement levels to ensure player risk detection systems can be
assessed across varying engagement levels.
Gambling Vertical refers to the type of gambling products offered. The main gambling
verticals by market share are lottery, casino, and sports betting. Within these verticals there
are sub-categories—for example, table games, slots, and live dealer in casino, draw products and instant win games in lottery, and fixed odds, prop, and parlay in sports betting.
While ML models can be developed to encapsulate player activity across gambling verticals, prior work has suggested superior performance when vertical-specific models are
developed (e.g., see Kairouz et al., 2023). Thus, a benchmarking suite should strive to provide datasets that can be used to develop both broad and vertical-specific models.

Example Benchmark Datasets
To illustrate the capabilities and practicality of the benchmarking suite concept, we present
four possible benchmark datasets that may be provided (Fig. 2). These examples are not

Fig. 2 Examples of benchmark datasets

intended to be exhaustive; rather, we offer them as a conceptual starting point, representing
some of the most common or likely scenarios. Each example demonstrates how the framework’s dimensions can provide different, measurable evaluation scenarios.
For each of these examples we make the assumption that each benchmark dataset consists
of players’ gambling data (i.e., session and/or transaction level information), is sufficient in
size, and has the required labels to build ML and predictive models (e.g., players labeled as
at-risk or not). We detail specific data requirements later in this section. Additionally, while
acknowledging the presence of other verticals, our examples and illustrations cover the top
three verticals (i.e., lottery, casino, sports betting).
Importantly, a single benchmark dataset can be used to support multiple benchmark tasks.
A task is defined as the specific ML exercise a participant must complete. For instance, a
dataset containing PGSI scores as the target variable could support multiple tasks. These
might include a binary classification problem, where the goal is predicting whether a player
is at risk based on PGSI scores (e.g., 0 = no risk, 5+ = at-risk), or a multi-class classification

problem, where the goal is predicting risk across multiple PGSI buckets (e.g., 0, 1-2, 3-4,
5-7, 8+). However, acknowledging that validated self-report measures are not always
available, tasks could also be defined around other harm proxies (See Section “Defining the
Target Variable” for a discussion).
A. Universal Benchmark Dataset: If an operator is multi-vertical (i.e., offers a wide
range of gambling products), and has built one risk detection algorithm to be used across
its entire player database, that operator can test that algorithm on a benchmark dataset that
encompasses player activity across each of the core dimensions. In other words, this labeled
dataset would contain session-level activity records that can be aggregated to longer time
horizons for players with various engagement levels across a range of gambling products,
providing variability on the Time, Engagement Level, and Gambling Vertical dimensions.
B. Lottery Players Benchmark Dataset: Some operators focus on offering a particular
gambling vertical, for example lottery, and therefore would benefit from a benchmark on
this subset of players. Additionally, evidence suggests that ML models for player risk detection may work best when they are tuned for specific activities (Ghaharian et al., 2023;
Kairouz et al., 2023).
C. Highly Engaged Players Benchmark Dataset: Data scientists typically focus on
sub-populations where data is more abundant and build models around these populations,
therefore a useful benchmark for industry could be more highly engaged players who have
reasonably long tenure (e.g., months rather than days).
D. Early Detection Benchmark Dataset: It is critical to detect risk as early as possible
to enable timely interventions and prevent future harm. Therefore, this benchmark dataset
would specifically assess an algorithm’s ability to predict future harm using only short-term
data. It would use data from a player’s initial days of activity (e.g., days 1-7) to predict a
negative outcome much further in the future (e.g., a risk flag at 6 months). Unlike the previous benchmark (C), this assessment could cover all engagement levels, not just the highlyengaged subset.
While the preceding examples (B–D) focus on isolating a single core dimension, realworld benchmark datasets will likely cross multiple dimensions simultaneously, which may
better reflect the complex nature of player risk detection in practice. For instance, detecting risky player sessions (e.g., those involving frequent top-ups or failed deposits) on slot
products would require a dataset spanning two dimensions: Gambling Vertical (e.g., slots)
and Time (sessions). Similarly, a common and valuable benchmark for a casino operator
might combine the Gambling Vertical and Engagement Level dimensions, specifically to
assess algorithms on more highly engaged, high-value players. Ultimately, these types of
composite benchmark datasets are vital for assessing algorithms under conditions that are
representative of the actual operational environment.

Requirements
As Dueben et al. (2022) describe, “Benchmark datasets are defined by the community.
While many datasets can be proclaimed as ‘benchmarks,’ they are only successful if
the ML community embraces them.” Similarly, the success of a benchmarking suite for
player risk detection hinges on its adoption by relevant stakeholders. As a foundation
for future implementation of such a benchmarking suite, we present initial discussion on
data specifications as well as various technical and usability requirements. Many of the

administrative requirements could be satisfied by existing data repositories (for in-depth
discussion see: Longjohn et al., 2024) and could be worked into a customized repository
specific for this task domain.

Data Collection
Data collection involves the curation of data into the benchmarking suite. It is unlikely
that the breadth of global stakeholders in the gambling sector will collectively cooperate to
curate a single, comprehensive dataset representing the entire gambling industry. Likewise,
the practicality, in terms of complexity and cost, of curating such a global dataset is likely
to be significant. A more realistic scenario involves gradual curation and procurement of
data from various stakeholders over time. In the past, industry stakeholders have donated
data to academic institutions (e.g., The Transparency Project and UNLV Payments Research
Collaborative), regulators have required licensees to transfer data to centralized storage for
audit purposes10, and there are examples of industry consortia supporting efforts to share
data across operators11.

Dataset Documentation
As data are curated, they must be presented in a way that allows stakeholders (including
academics, researchers, and industry operators) to effectively use these datasets for evaluating ML models.
Because raw data are rarely ready for immediate use, they require processing to become a
usable benchmark dataset. The benchmarking suite must facilitate recording of any processing steps taken, and given the sensitivity of player protection – where model performance
directly impacts human well-being – there should also be an attempt to enforce quality
review. As highlighted by Longjohn et al. (2024), there is currently a lack of quality review
of datasets, and existing repositories often view themselves as simply providing infrastructure to support data sharing rather than providing review or quality assurance processes.
To address this, the proposed benchmarking suite could require the publication of peerreviewed introductory papers to accompany datasets, similar to those found in Data in Brief
(e.g., see: Ghaharian et al., 2023).
The entire suite would house numerous datasets. Thus, users must understand each
dataset’s context and structure. Best practices from established data science platforms and
repositories can provide a strong foundation (e.g., HuggingFace, UCI ML Repository,
OpenML). For instance, datasets should include core metadata such as the size and shape
of the data, list a creator/owner of the dataset (who can act as a point of contact for queries
and/or issues), as well as versioning and timestamps. Some datasets may well reside in
raw format (e.g., session-level activity for each bet placed), whereas others may represent
aggregated information (e.g., session-level activity rolled up to daily or weekly information

For example, the Netherlands Gambling Authority (Kansspelautoriteit, Ksa) mandates the use of a Controle
Data Bank (CDB).

For example, the Responsible Online Gaming Association (ROGA) has announced the creation of a data
clearing house that will consolidate and share player self-exclusion data among its members to advance consumer protection. Similarly, the GamProtect scheme facilitated by the Betting and Gaming Council seeks to
create a mechanism for participating operators to compliantly and securely share information with each other.

about players’ activity). Critically, documentation should specify the temporal relationship
between the player activity data and any associated labels (the target variable). This includes
defining the timeframe the activity data covers (e.g., the first 7 days of play, 12-months
since registration date) and its temporal proximity to the target variable (e.g., a PGSI score
collected at 6 months, the date self-exclusion occurred). Tasks associated with benchmark
datasets could also be listed alongside this metadata and/or with further details about each
task (e.g., train-test splits, primary evaluation metric, etc.) provided in a separate location
within the repository (e.g., as a downloadable file or a link to a dedicated page). We provide
an example of such metadata for the Early Detection Benchmark Dataset (example D above)
in Table 1.
Table 1 Example metadata for a
benchmark dataset

Attribute
Dataset name
Description

Dimensions
Benchmark
task(s)

Target
variable
Size
Data fields

Data
dictionary
Creator
Author
Citation
Versioning

Description
The Matrix Casino: Early Risk Dataset (Zion)
Contains session-level data from The Matrix online
casino (Zion, Earth). The dataset includes 105,677
sessions from 5,465 unique players during their first
7 days of activity.
Time: 7 days
Engagement level: All players who completed PGSI
Vertical: Casino
B1 7 day detection: Supervised Classification. To
predict a player’s future harm status using only their
initial 7 days of session data.
B2 1 day detection: Supervised Classification. To
predict a player’s future harm status using only their
initial day of session data.
Risk_Flag: A binary flag (0 or 1) indicating high-risk
status at 6 months, defined by a PGSI score of 5+.
105,677 rows (sessions)
Player_ID: Unique player identifier.
Transaction_Type: A deposit or withdrawal of real
money.
Transaction_Amount: Amount of real money
deposited or withdrawn.
Transaction_Method: Method of payment: card,
PayPal etc.
Transaction_Status: Status of the transaction:
approved, declined, etc.
Start_Time: Session start timestamp (UTC).
End_Time: Session end timestamp (UTC).
Bet_Count: Number of wagers in the session.
Total_Staked: Total monetary value staked.
Net_Outcome: Net win/loss from the session.
Product: Gambling product used.
Risk_Flag: (Target) Player’s risk status at 6 months.
Download_link
Mr. Thomas A. Anderson
Anderson, T. A. (2025). The Matrix Casino: Early
Risk Dataset (Zion) https://doi.org/00.0000/prd.
bs.2025.1117
V.1 2025-11-17

This “data work” is a substantial undertaking but is essential to the downstream success
of any benchmark dataset. Unfortunately, this type of work is often undervalued and underincentivized (Gero et al., 2023). To incentivize and recognize these efforts, datasets could
be treated as citable intellectual contributions through persistent identifiers, such as a Digital
Object Identifier (DOI) (Longjohn et al., 2024).

Dataset Access and Usability
Datasets and associated benchmark tasks should be easy to access and use. Datasets should
be provided in common, accessible formats (e.g., CSV, various application programming
interfaces [API] formats) with detailed documentation on how to access data.
For tasks, a baseline model or script (e.g., logistic regression or random forest) could be
provided in a popular language (e.g., Python). This serves as a reference implementation,
demonstrating how to load the data and generate predictions. It also establishes a performance baseline that more sophisticated models can attempt to surpass, offering an immediate point of comparison.

Result Submission and Validation
Benchmark tasks will also require a trustworthy mechanism for users to submit their model’s results and have them validated. To ensure fair comparisons, users must adhere to a
clear set of rules. There are several potential approaches for managing this process, each
with its own trade-offs, we outline these in Table 2.

Challenges and Considerations
While our framework in Section “A Proposed Benchmarking Framework for Player Risk
Detection Systems” provides a foundational structure, it is intended as a conceptual guide.
We recognize that the practical implementation of such a suite faces significant hurdles and
limitations not fully addressed in this initial proposal. This section, therefore, discusses several key challenges and practical considerations that our field must navigate to move from
this conceptual model to a functional, robust benchmarking system.

Governance and Administration
A central question in implementing such a framework is: who should govern and administer the system? This governing entity would be responsible for overseeing all functions
including data collection, administration of contributors and users, ensuring correct documentation, facilitating usability, and validating the robustness and trustworthiness of the
benchmark process and its results.
Administration by a governmental body is likely to improve coordination by operators
and overcome collective action problems inherent in a competitive private market. Regulators can legally mandate the submission of standardized data from all licensed operators,
ensuring the comprehensiveness and consistency required for a foundational dataset.

Table 2 Result submission and
validation options for benchmark
tasks

Method
Publicationbased
verification
Full code
and model
submission

Prediction
submission

Container
submission

Description

Strengths and
limitations
The solution can act as a data
Rigorous
repository (e.g., UCI Machine academic peer
Learning Repository) and
review but
results are validated indirectly slow and not
via academic peer review.
standardized.
Participants submit complete
This is maxisource code, configuration,
mum reproand model details along with
ducibility, but
their results, enabling the
conflicts with
benchmark owners to directly potential IP
reproduce outcomes.
concerns.
Akin to Kaggle, users run
This addresses
models locally on the benchIP concerns, but
mark dataset, then submit
cannot know
only the predictions which are how predictions
scored against hidden labels.
were generated.
Users submit an executable
Addresses not
model (e.g., via a Docker
knowing how
container) and the benchmark predictions were
owners run the model against a generated, but
hidden test set.
requires more
engineering and
resources.

Note: The tasks could potentially support all submission pathways and
implement a “badge system” to recognize different levels of validation
(e.g., bronze for prediction-only, silver for containerized submission,
gold for full code + publication)

A third-party model involving a neutral entity like an international research institute or
data trust presents a distinct advantage in this regard (Paprica et al., 2020). Such an organization is not constrained by jurisdictional borders and could therefore aggregate data across
multiple states, provinces, and nations. This transnational approach would enable a far more
comprehensive dataset.

Dataset Size
A robust benchmark dataset must address sample considerations that extend beyond conventional calculations of statistical power. These include challenges with representativeness
and detecting low-prevalence phenomena. For instance, problematic gambling, as identified
by a proxy like the PGSI, has a low base rate, necessitating a very large overall sample
just to capture enough positive cases for reliable model training and evaluation. Without
an adequate number of these cases, a model’s ability to accurately identify at-risk players
becomes statistically unreliable.
Furthermore, the database must accommodate a wide range of models, from simple
logistic regression to complex deep learning networks. Since more complex models require
larger datasets to prove their effectiveness, sample sizes should be large enough for the most
sophisticated systems. While sourcing such a dataset could prove challenging, potential
solutions include the construction of synthetic datasets, government-mandated data aggregation, or the pooling of data through a third party.

The use of synthetic data provides a potentially constructive path to scale dataset size
while avoiding issues with player privacy and commercial data sharing sensitivity (Samsel
et al., 2025). By generating an artificial dataset that replicates the statistical properties of
real-world data without containing any personally identifiable information, synthetic data
allows for broad access for research and model testing while the source data remains secure.
The viability of this approach, however, is entirely contingent upon the quality and representativeness of the original dataset from which the synthetic version is derived, as well
as the assumptions of the approach used to generate the synthetic data (Chen et al., 2021;
Susser et al., 2024).

Defining the Target Variable
A review of the academic literature and existing solutions reveals at least three different
categories of outcome targets for ML models in player risk detection.
First, proxy indicators of harm such as account closure and, more commonly, voluntary
self-exclusion (VSE) are used by both industry and academic studies (e.g., Finkenwirth
et al., 2021; Hopfgartner et al., 2023; Percy et al., 2016). These are routinely collected
alongside the data used to make predictions, can easily delineate the at-risk group through
binary classification (VSE or no VSE), and are not subject to self-report bias or measurement error. However, their reliability as indicators of at-risk status is questionable. Players
can choose to self-exclude or close their accounts for various reasons beyond the experience
of gambling problems, including to pre-empt harms or to close one’s online account (Auer
& Griffiths, 2016; Catania & Griffiths, 2021). One strategy to mitigate this risk is to use a
subset of self-excluders who have a sufficiently long track record of play with the operator,
to ensure players who self-exclude relatively quickly after opening an account are not used
for model training.
Second, screening instruments such as the PGSI (Ferris & Wynne, 2001) or Brief Biosocial Gambling Screen (Gebauer et al., 2010) have been used as target variables in recent academic studies (Auer & Griffiths, 2023; Hopfgartner et al., 2024; Kairouz et al., 2023; Murch
et al., 2023; Perrot et al., 2022) and by Focal Research Consultants (Hancock et al., 2008;
Schellinck & Schrans, 2011). Self-report measures offer a more reliable indicator of risk
status, although are more difficult to obtain and link with the data used to make predictions,
are subject to self-report biases (particularly if administered by the operator), and require a
threshold for classification as at-risk to be selected in advance (Murch et al., 2023). Additionally, the measures reflect a respondent’s overall gambling behavior, which may span
multiple platforms, operators, or even jurisdictions, limiting its interpretability as a label for
data from any single operator’s platform. Furthermore, because self-reported risk status can
change over time in response to variations in gambling engagement, repeated administration
of screening instruments would improve datasets. Longitudinal measurement of the target
variable would allow models to be trained and evaluated on their ability to predict transitions into or out of at-risk status, rather than relying solely on cross-sectional classification.
Third, a final approach may be the use of single indicators such as loss chasing or
substantial variations from normal play, which are themselves sometimes used to classify
consumers as at-risk in the absence of ML (Delfabbro et al., 2024; McAuliffe et al., 2022).
Several researchers have attempted to develop thresholds for standard behavioral indicators
(e.g., frequency of gambling, amount spent) that differentiate lower- from higher-risk

gambling to form lower risk gambling limits (Hodgins et al., 2023; Louderback et al.,
2021) – these, or similar threshold-style indicators, could plausibly serve as targets, either
individually or in combination.
Overall, the variation and lack of a standard approach to classification presents a challenge to constructing and comparing ML models, as there is no agreed-upon specification
to label training data appropriately. Given the diversity of outcome variables and definitions of risk, we do not propose a single ground truth. Instead, we recommend a suite of
benchmarks that can evolve over time as the domain continues to develop and as a clearer
definition of “at-risk” emerges. We do, however, advocate for transparency in how at-risk
status is determined to ensure the process can be replicated in real-world environments and
to help acknowledge the biases inherent in different approaches. For example, human-based
labeling of cases based on opaque reasoning would not satisfy this criterion, but labels
determined using a clear and reproducible set of steps would.

Demographic and Jurisdictional Diversity
Another challenge is significant diversity in player populations and jurisdictions. A successful benchmarking framework must avoid over-representing datasets from one demographic
group (e.g., all men) or market (e.g., North America), except where the benchmark results
are explicitly only applied within those settings. Over-representing certain demographic
groups in benchmark datasets risks validating the performance of models that perform
poorly in real-world settings with more diverse populations and could potentially favor
biased models that disproportionately target or ignore minority groups (see Murch et al.,
2024a).
There is evidence that ML models can effectively detect at-risk gambling in datasets
from countries not included in training data, although training on local data can improve
performance (Hopfgartner et al., 2024). A fairer benchmarking framework will offer datasets from a variety of regions to allow contributors to demonstrate the performance of their
model(s) in preferred settings (e.g., EU vs. North America).

Dataset Characteristics
Jurisdictional variation may also impact the ease with which models are applied due to
differences in feature availability. Additionally, differences in available products and other
policies across jurisdictions will affect model applicability. This may skew the composition
and relevance of benchmark datasets developed in those contexts. For example, a model
built using online tracking data in Australia where players can cancel their withdrawals
may attribute high importance to this behavior, which would not be present in datasets from
jurisdictions that have prohibited this feature (e.g., Great Britain; UK Gambling Commission, 2021). This highlights the need for clear documentation that describes the characteristics of each benchmarking dataset and idiosyncrasies resulting from regulation or product
innovation.
Those experienced in working with tracking data will be familiar with variations in labeling (e.g., “stake” vs. “amount”; “bonus stake” vs. “free stake”) and the operationalization
of variables (e.g., winning bets recorded as negative vs. positive values; aggregation or
separation of multi-bet legs) between datasets from different operators. While minor, these

discrepancies could preclude the efficient application of models. For benchmarking to work,
a harmonized data dictionary and standardized preprocessing steps will be essential to produce benchmark datasets that others can replicate in model training.
A further consideration is the temporal structure of the data. Academic studies report
building models using varying timeframes of tracking data, ranging from 12 months (e.g.,
Murch et al., 2023; Perrot et al., 2022) to just 30 days (Hopfgartner et al., 2024). As discussed in Section “A Proposed Benchmarking Framework for Player Risk Detection Systems”, there is a need to provide benchmarks for different periods that span months to days.
Ideally, benchmark datasets will span an extended period (e.g., 12 months) to permit varied
testing on the same dataset and the use of diverse timeframe aggregations (e.g., monthly
or daily behavioral averages). A benchmark for acute risk detection (e.g., 1 to 7 days) may
benefit from uniquely large datasets, since these can include all active and newly registered
customers, not only those with extensive historical data.

The Evolving Gambling Landscape and need for Dynamic Benchmarks
A final challenge is the dynamic and evolving nature of gambling. Models built to accurately classify or predict stable outcomes such as handwriting or images can be continuously
evaluated against benchmark datasets with little need for updating. By contrast, gambling
environments and consumer behavior can be continuously impacted by changes in regulation, local economies, and product innovation, which may result in data drift. Validation of
a model against a historical dataset (such as those added to The Transparency Project) may
hold little weight when evaluating contemporary performance. As such, ML benchmarks for
gambling cannot be static. Government mandates that require data sharing (see Newall and
Swanton, 2024) are likely to be the most successful way of accessing periodically updated
datasets for a benchmarking suite. Regardless, successfully sharing and deploying the first
iteration of benchmarking datasets without compromising privacy will be essential to building trust with operators and the public alike, and ensuring continuous buy-in from industry
stakeholders.

Conclusion
Technology, specifically AI and its subsets such as ML, are being increasingly used to support industry efforts to minimize harms. While we applaud efforts to leverage technology
to better protect consumers, the current state of affairs is sub-optimal, in that it is not possible to credibly ascertain the efficacy of solutions that use ML to detect at-risk play. As AI
becomes increasingly integrated into diverse processes, ranging from business to health
and education, benchmarks are becoming increasingly developed to enable more objective
assessment and comparison of their efficacy for the task at hand.
The gambling industry has an opportunity to leverage these technological developments
and lead in responsible and ethical AI by promoting greater transparency and trust in the
harm detection solutions they develop and deploy. Benchmarking is key to this, and there
are significant examples across other use cases and industries on how benchmarking can be
developed successfully.

The concept of benchmarking is particularly important in any gambling industry practice
that relies on ML models to support decision making in the areas of consumer protection,
fairness, and crime. For example, these concepts could be applied to assessing the efficacy
of sports betting integrity (i.e., match-fixing) detection algorithms, anti-money laundering
(AML) detection algorithms, and impacts of operator customer RG interventions.
While we outline a broad range of challenges and considerations, we argue that these are
not insurmountable and can be addressed with the right combination of data access, domain
and technical expertise, and the ability to develop trustworthy solutions. A failure of industry to address these concerns will cast doubt on the efficacy of these solutions and increase
pressure on regulators to take greater control of how they are developed and used. In Spain,
the Directorate General for the Regulation of Gambling (DGOJ) recently received a legal
mandate to establish a unified ML model that will be mandatory for all licensed operators in
Spain (Directorate General for the Regulation of Gambling, 2025). Such regulatory developments are likely to continue expanding if the transparency of existing industry methods
remains opaque.
Acknowledgements Kasra Ghaharian would like to thank Piyush Puranik for his useful comments regarding
the result submission and validation options for benchmark tasks.
Author Contributions Kasra Ghaharian: Conceptualization, Writing – Original Draft, Writing – Review
& Editing, Visualization, Supervision, Project administration. Simo Dragicevic: Conceptualization, Writing
– Original Draft, Writing – Review & Editing, Visualization, Project administration. Chris Percy: Conceptualization, Writing – Original Draft, Writing – Review & Editing. Sarah E. Nelson: Writing – Original Draft,
Writing – Review & Editing. W. Spencer Murch: Writing – Original Draft, Writing – Review & Editing.
Robert M. Heirene: Writing – Original Draft, Writing – Review & Editing. Kahlil Simeon-Rose: Writing –
Original Draft, Writing – Review & Editing. Tracy Schrans: Writing – Review & Editing.
Funding No funding was received for this work. The authors have stated their relevant disclosures in the
Disclosures section below.

Declarations
Competing Interests The authors have stated their competing interests and disclosures in the Disclosures
section below.
Ethics approval and consent to participate This manuscript is a commentary and does not involve the collection or analysis of primary data from human participants or animals.
Disclosures During the last five years, the International Gaming Institute (IGI) at University of Nevada,
Las Vegas, has received funding for its research and programs from Action Gaming, American Gaming
Association, Aristocrat Leisure Limited, Association of Gaming Equipment Manufacturers, Axes.ai, Bet
Blocker, Clarion Gaming, DraftKings, Entain Foundation, ESPN, Evoke plc, Focal Research Consultants,
Gaming Analytics, Global Payments, IGT, Kindbridge Behavioral Health, Las Vegas Sands Corporation,
Massachusetts Gaming Commission, MGM Resorts International, Playtech plc, Responsible Online Gaming Association, Yuhaaviatam of San Manuel Nation, Sightline Payments, Sports Betting Alliance, State
of Nevada Department of Health and Human Services, State of Nevada Knowledge Fund, Walker Digital
Table Systems, and Wynn Resorts Ltd. Additionally, IGI organizes the triennial International Conference on
Gambling and Risk Taking, a research-focused event supported by sponsors from industry, academia, and
the legal/regulatory sectors; a full list of sponsors for the most recent conference can be found at ​h​t​t​p​s​:​​/​/​w​w​
w​​.​u​n​l​v​.​​e​d​u​/​​i​g​i​/​c​​o​n​f​e​r​​e​n​c​e​/​1​​8​t​h​/​​s​p​o​n​s​o​r​s. IGI is home to an industry-focused advisory board ​(​h​t​t​p​s​:​/​/​w​w​w​.​u​
n​l​v​.​e​d​u​/​i​g​i​/​a​d​v​i​s​o​r​y​-​b​o​a​r​d​​​​)​, and specific programs, such as AiR Hub, have their own advisory panels. These
advisory roles include resource support, and individual advisors are required to adhere to IGI policies. IGI
maintains a strict research policy (https://www.unlv.edu/igi/research-policy), as well as a partnership and

transparency framework (​h​t​t​p​s​:​​/​/​w​w​w​​.​u​n​l​v​.​​e​d​u​/​​i​g​i​/​p​​o​l​i​c​i​​e​s​/​p​a​r​​t​n​e​r​​s​h​i​p), to ensure appropriate firewalls exist
between funding entities and IGI’s research and programs.
During the past 5 years, Kasra Ghaharian has received funding for research and/or consulting services from
the Nevada Department of Health and Human Services, the Nevada Governor’s Office of Economic Development, the Massachusetts Gaming Commission, AXES.ai, Playtech, Sightline, IGT, Differential, Focal
Research Consultants, GP Consulting, and the International Center for Responsible Gaming. Ghaharian has
received honoraria/travel reimbursement from the Responsible Gambling Council, the Illinois Council on
Problem Gambling, and Kindred Group. None of these entities played roles in the design, analysis, or interpretation of research, and imposed no constraints on publishing.
During the past 5 years, Simo Dragicevic has received funding for employment, research, and consulting
services from The European Lotteries, The Gambling Commission (Great Britain), Playtech Plc, and Yaspa.
Simo is also Adjunct Fellow at the International Gaming Institute (IGI).
Chris Percy provides consulting services to the gambling industry, with recent activity including research for
the UK Gambling Commission and the BetBuddy player risk detection software and responsible gambling
R&D at Playtech Plc.
During the past five years, Sarah E. Nelson has served as a paid grant reviewer for the New Frontiers in
Research Fund (Government of Canada), the International Center for Responsible Gaming (ICRG), and the
National Institutes of Health (NIH). She has also received travel reimbursement and speaker honoraria from
the ICRG, Responsible Gaming Association of New Mexico, Responsibility.org, and the Maryland Highway Safety Office. She received publication royalty fees from the American Psychological Association, and
received course royalty fees from the Harvard Medical School Department of Continuing Education.
W. Spencer Murch holds a postdoctoral fellowship from the Alberta Gambling Research Institute, a joint
initiative by the Alberta Provincial Government, and the Universities of Calgary, Lethbridge, and Alberta.
WSM holds a research grant from the International Centre for Responsible Gaming (USA) that did not fund
his contributions to this research. WSM served on the Advisory Board on Safer Gambling, an expert group
that formerly advised the UK Gambling Commission.
Robert M. Heirene’s contribution to this work was supported by a post-doctoral fellowship from the New
South Wales Responsible Gambling Fund as part of the Gambling Research Capacity Grant program, administered by the Office of Responsible Gambling. RMH has worked on a project funded by Responsible Wagering Australia (a representative body of Australian online wagering operators; University of Sydney, 2019–
2021) and as an independent, sub-contracted statistical consultant for PRET Solutions Inc on a commissioned
project (funded by the Australian Casino operator Crown; 2023). RMH has also received research funding
from the International Centre for Responsible Gaming (ICRG) and the Brain and Mind Center, University
of Sydney.
Khalil Simeon-Rose declares the following financial interests/personal relationships which may be considered as potential competing interests: In the three-year period from 2022 to 2025, KS received research
funding from the Cambridge Health Alliance, Division on Addiction; he received consulting payments from
Eilers & Krejcik Gaming, Scientific Affairs, VictorStrategies, MGM Resorts, and the United Arab Emirates’
General Commercial Gaming Regulatory Authority; and he received expert witness payments from Benjamin Leigh Carter. KS holds unpaid appointments on the advisory board of the Responsible Online Gaming
Association, the Canadian Gaming Association’s responsible gaming sub-committee, and the Deutsche Stiftung Glücksspielforschung gGmbH Scientific Advisory Board.
Tracy Schrans declares that, during the past 5 years, Focal Research Consultants have received funding from
various gaming companies, regulatory bodies, government funding agencies, and research institutions for
consultancy work in the area of player data analytics, social responsibility, and safer gambling products, services, and compliance solutions. None of these entities played a role in the design, analysis, or interpretation
of this study, and imposed no constraints on publishing.
Consent for publication Not applicable.

References
Adami, N., Benini, S., Boschetti, A., Canini, L., Maione, F., & Temporin, M. (2013). Markers of unsustainable gambling for early detection of at-risk online gamblers. International Gambling Studies, 13,
188–204. https://doi.org/10.1080/14459795.2012.754919
Andrews, B. K., Linden, M. V., & Wohl, M. J. A. (2025). A framework for independent research using industry funding: the massachusetts model. International Gambling Studies, 1–12. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​0​/​1​
4​​4​5​9​7​9​​5​.​2​0​2​5​​.​2​5​0​​8​4​6​0

Auer, M., Griffiths, M., 2016. Should voluntary" self-exclusion" by gamblers be used as a proxy measure for
problem gambling? MOJ Addiction Medicine & Therapy 2, 00019. 10.15406/mojamt.2016.2.00019.
Auer, M., & Griffiths, M. D. (2023). Using artificial intelligence algorithms to predict self-reported problem
gambling with account-based player data in an online casino setting. Journal of Gambling Studies, 39,
1273–1294. https://doi.org/10.1007/s10899-022-10139-1
Auer, M., & Griffiths, M. D. (2024). Predicting high-risk gambling based on the first seven days of gambling
activity after registration using account-based tracking data. International Journal of Mental Health and
Addiction, 22, 3397–3413. https://doi.org/10.1007/s11469-023-01056-4
Backlund, A., Petersson, L., 2025. Vending-bench: A benchmark for long-term coherence of autonomous
agents. arXiv:2502.15840.
Bean, A.M., Kearns, R.O., Romanou, A., Hafner, F.S., Mayne, H., Batzner, J., Foroutan, N., Schmitz, C.,
Korgul, K., Batra, H., et al., 2025. Measuring what matters: Construct validity in large language model
benchmarks. arXiv:2511.04703.
Bedford, K., & Chandan, J. S. (2024). High stakes: does the gambling white paper go far enough? The Lancet, 403, 1236–1237.
Binesh, F., Ghaharian, K., 2025. Identifying risks and ethical considerations of ai in gambling: A scoping review.
International Journal of Hospitality & Tourism Administration, 1–31 10.1080/15256480.2025.2494575.
Biomarkers Definitions Working Group. (2001). Biomarkers and surrogate endpoints: preferred definitions
and conceptual framework. Clinical pharmacology & therapeutics,69, 89–95. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​.​1​0​6​7​/​
m​c​p​.​2​0​0​1​.​1​1​3​9​8​9​​​​
Braverman, J., & Shaffer, H. J. (2012). How do gamblers start gambling: Identifying behavioural markers
for high-risk internet gambling. The European Journal of Public Health, 22, 273–278. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​
1​0​.​1​0​9​3​/​e​u​r​p​u​b​/​c​k​p​2​3​2​​​​
Broda, C., Limao, N., & Weinstein, D. E. (2008). Optimal tariffs and market power: the evidence. American
Economic Review, 98, 2032–2065. https://doi.org/10.1257/aer.98.5.2032
Catania, M., & Griffiths, M. D. (2021). Understanding online voluntary self-exclusion in gambling: An
empirical study using account-based behavioral tracking data. International Journal of Environmental
Research and Public Health, 18, 2000. https://doi.org/10.3390/ijerph18042000
Chen, R. J., Lu, M. Y., Chen, T. Y., Williamson, D. F. K., & Mahmood, F. (2021). Synthetic data in machine
learning for medicine and healthcare. Nature Biomedical Engineering, 5, 493–497. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​.​1​
0​3​8​/​s​4​1​5​5​1​-​0​2​1​-​0​0​7​5​1​-​8​​​
Collins, G.S., Moons, K.G.M., Dhiman, P., Riley, R.D., Beam, A.L., Van Calster, B., Ghassemi, M., Liu, X.,
Reitsma, J.B., van Smeden, M., Boulesteix, A.L., Camaradou, J.C., Celi, L.A., Denaxas, S., Denniston,
A.K., Glocker, B., Golub, R.M., Harvey, H., Heinze, G., Hoffman, M.M., Kengne, A.P., Lam, E., Lee,
N., Loder, E.W., Maier-Hein, L., Mateen, B.A., McCradden, M.D., Oakden-Rayner, L., Ordish, J., Parnell, R., Rose, S., Singh, K., Wynants, L., Logullo, P., 2024. Tripod+ai statement: updated guidance for
reporting clinical prediction models that use regression or machine learning methods. BMJ, e078378
10.1136/bmj-2023-078378.
Cruz Rivera, S., Liu, X., Chan, A. W., Denniston, A. K., Calvert, M. J., Ashrafian, H., Beam, A. L., Collins,
G. S., Darzi, A., Deeks, J. J., ElZarrad, M. K., Espinoza, C., Esteva, A., Faes, L., Ferrante di Ruffano, L.,
Fletcher, J., Golub, R., Harvey, H., Haug, C., … Yau, C. (2020). Guidelines for clinical trial protocols
for interventions involving artificial intelligence: the spirit-ai extension. The Lancet Digital Health, 2,
e549–e560. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​1​6​/​s​2​5​8​9​-​7​5​0​0​(​2​0​)​3​0​2​1​9​-​3
Delfabbro, P., Parke, J., & Catania, M. (2023). Behavioural tracking and profiling studies involving objective data derived from online operators: A review of the evidence. Journal of Gambling Studies, 40,
639–671. https://doi.org/10.1007/s10899-023-10247-6
Delfabbro, P., Parke, J., Catania, M., Chikh, K., 2024. Behavioural Markers of Harm and Their Potential in
Identifying Product Risk in Online Gambling. International Journal of Mental Health and Addiction
22, 3451–3469. ​h​t​t​p​s​:​​/​/​l​i​n​​k​.​s​p​r​i​​n​g​e​r​​.​c​o​m​/​​1​0​.​1​0​​0​7​/​s​1​1​​4​6​9​-​​0​2​3​-​0​1​0​6​0​-​8, DOI: ​1​0​.​1​0​0​7​/​s​1​1​4​6​9​-​0​2​3​-​0​1​
0​6​0​-​8​.​​​
Directorate General for the Regulation of Gambling, 2025. Regulation of gambling. ​h​t​t​p​s​:​​/​/​w​w​w​​.​b​o​e​.​e​​s​/​b​u​​s​
c​a​r​/​​d​o​c​.​p​​h​p​?​i​d​= ​​B​O​E​-​​A​-​2​0​2​3​-​6​7​3​5. Accessed: 2025–11-25.
Dueben, P. D., Schultz, M. G., Chantry, M., Gagne, D. J., Hall, D. M., & McGovern, A. (2022). Challenges
and benchmark datasets for machine learning in the atmospheric sciences: Definition, status, and outlook. Artificial Intelligence for the Earth Systems, 1, Article e210002. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​.​1​1​7​5​/​A​I​E​S​-​D​-​2​
1​-​0​0​0​2​.​1​​​
Entain, 2022. Entain Sustain ESG showcase event. ​h​t​t​p​s​:​​/​/​w​w​w​​.​e​n​t​a​i​​n​g​r​o​​u​p​.​c​o​​m​/​n​e​w​​s​-​i​n​s​i​​g​h​t​s​​/​l​a​t​e​​s​t​-​n​e​​w​s​/​
2​0​2​​2​/​e​n​​t​a​i​n​-​​s​u​s​t​a​​i​n​-​e​s​g​​-​s​h​o​​w​c​a​s​e​-​e​v​e​n​t​/. 19 Oct 2022.
von Eschenbach, W. J. (2021). Transparency and the black box problem: Why we do not trust ai. Philosophy
& Technology, 34, 1607–1622. https://doi.org/10.1007/s13347-021-00477-0

Ferris, J. A., & Wynne, H. J. (2001). The Canadian problem gambling index. ON: Canadian Centre on substance abuse Ottawa.
Finkenwirth, S., MacDonald, K., Deng, X., Lesch, T., & Clark, L. (2021). Using machine learning to predict
self-exclusion status in online gamblers on the playnow. com platform in british columbia. International
Gambling Studies, 21, 220–237. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​0​/​1​4​​4​5​9​7​9​​5​.​2​0​2​0​​.​1​8​3​​2​1​3​2
Forrest, D., McHale, I.G., 2022. Patterns of Play Technical Report 2: Account Data Stage. Technical Report.
NatCen Social Research. ​h​t​t​p​s​:​​/​/​w​w​w​​.​n​a​t​c​e​​n​.​a​c​​.​u​k​/​s​​i​t​e​s​/​​d​e​f​a​u​l​​t​/​f​i​​l​e​s​/​2​​0​2​3​-​0​​3​/​P​a​t​t​​e​r​n​s​​%​2​0​o​f​​%​2​0​P​l​​a​y​
_​T​e​c​​h​n​i​c​​a​l​%​2​0​​R​e​p​o​r​​t​%​2​0​2​_​​A​c​c​o​​u​n​t​%​2​0​D​a​t​a​%​2​0​S​t​a​g​e​%​2​0​R​e​p​o​r​t​.​p​d​f.
Gebauer, L., LaBrie, R., & Shaffer, H. J. (2010). Optimizing dsm-iv-tr classification accuracy: A brief biosocial screen for detecting current gambling disorders among gamblers in the general household population. The Canadian Journal of Psychiatry, 55, 82–90. https://doi.org/10.1177/070674371005500204
Gero, K. I., Das, P., Dognin, P., Padhi, I., Sattigeri, P., & Varshney, K. R. (2023). The incentive gap in data
work in the era of large models. Nature Machine Intelligence, 5, 565–567. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​.​1​0​3​8​/​s​4​2​
2​5​6​-​0​2​3​-​0​0​6​7​3​-​x​​​
Ghaharian, K., Abarbanel, B., Kraus, S. W., Singh, A., & Bernhard, B. (2024). Evaluating the generalizability
of payment behavioral profiles across gambling brands. International Gambling Studies, 24, 152–169. ​h​
t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​0​/​1​4​​4​5​9​7​9​​5​.​2​0​2​3​​.​2​2​1​​8​4​6​0
Ghaharian, K., Abarbanel, B., Phung, D., Puranik, P., Kraus, S., Feldman, A., & Bernhard, B. (2022). Applications of data science for responsible gambling: a scoping review. International Gambling Studies, 23,
289–312. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​0​/​1​4​​4​5​9​7​9​​5​.​2​0​2​2​​.​2​1​3​​5​7​5​3
Ghaharian, K., Binesh, F., Soligo, M., Golab, L., & Abarbanel, B. (2025a). AI ethics in a controversial industry: the case of gambling and its ethical paradox. AI and Ethics,5, 1685–1701. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​.​1​0​0​7​/​
s​4​3​6​8​1​-​0​2​4​-​0​0​5​2​0​-​8​​​
Ghaharian, K., Puranik, P., Abarbanel, B., Taghva, K., Kraus, S. W., Singh, A., Feldman, A., & Bernhard, B.
(2023). Payments transaction data from online casino players and online sports bettors. Data in Brief,
48, Article 109077. https://doi.org/10.1016/j.dib.2023.10907
Ghaharian, K., Soligo, M., Bischoff, J., 2025b. AI and player risk identification and response. Research
Report. University of Nevada, Las Vegas, International Gaming Institute. 10.17605/OSF.IO/W497Y.
Gomar, J. J., Bobes-Bascaran, M. T., Conejero-Goldberg, C., Davies, P., Goldberg, T. E., Initiative, A. D.
N., et al. (2011). Utility of combinations of biomarkers, cognitive markers, and risk factors to predict
conversion from mild cognitive impairment to alzheimer disease in patients in the alzheimer’s disease
neuroimaging initiative. Archives of general psychiatry, 68, 961–969. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​0​1​/​a​r​​c​h​g​e​n​​
p​s​y​c​h​i​​a​t​r​y​​.​2​0​1​1​.​9​6
Hancock, L., Schellinck, T., & Schrans, T. (2008). Gambling and corporate social responsibility (csr): Redefining industry and state roles on duty of care, host responsibility and risk management. Policy and
Society, 27, 55–68. https://doi.org/10.1016/j.polsoc.2008.07.005
Hanna, M., Pedeni, F., Suglia, A., Testoni, A., Bernardi, R., 2022. ACT-thor: A controlled benchmark for
embodied action understanding in simulated environments, in: Calzolari, N., Huang, C.R., Kim, H.,
Pustejovsky, J., Wanner, L., Choi, K.S., Ryu, P.M., Chen, H.H., Donatelli, L., Ji, H., Kurohashi, S.,
Paggio, P., Xue, N., Kim, S., Hahm, Y., He, Z., Lee, T.K., Santus, E., Bond, F., Na, S.H. (Eds.), Proceedings of the 29th International Conference on Computational Linguistics, International Committee on
Computational Linguistics, Gyeongju, Republic of Korea. pp. 5597–5612. ​h​t​t​p​s​:​/​/​a​c​l​a​n​t​h​o​l​o​g​y​.​o​r​g​/​2​0​
2​2​.​c​o​l​i​n​g​-​1​.​4​9​5​/​​​.​​​​
Hardy, A., Hardy, M., Kochenderfer, M., Lamparth, M., Reuel, A., Smith, C., 2024. Betterbench: Assessing
ai benchmarks, uncovering issues, and establishing best practices, in: Advances in Neural Information Processing Systems 37, Neural Information Processing Systems Foundation, Inc. (NeurIPS). p.
21763–21813. 10.52202/079017-0685.
Hastie, T., Tibshirani, R., Friedman, J., 2009. The Elements of Statistical Learning: Data Mining, Inference, and
Prediction. Springer Series in Statistics. 2nd ed., Springer, New York. DOI: 10.1007/978-0-387-84858-7.
Heirene, R.M., Wohl, M.J., 2025. Setting limits on expectations: rethinking who should be using responsible
gambling tools and why? International Gambling Studies, 1–8 10.1080/14459795.2025.2564402.
Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., Steinhardt, J., 2020. Measuring massive multitask language understanding. 10.48550/ARXIV.2009.03300.
Hodgins, D. C., Young, M. M., Currie, S. R., Abbott, M., Billi, R., Brunelle, N., Costes, J. M., Dufour, M.,
Flores-Pajot, M. C., Olason, D. T., et al. (2023). Lower-risk gambling limits: Linked analyses across
eight countries. International Gambling Studies, 23, 328–344. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​0​/​1​4​​4​5​9​7​9​​5​.​2​0​2​2​​
.​2​1​4​​3​5​4​6
Hopfgartner, N., Auer, M., Griffiths, M. D., & Helic, D. (2023). Predicting self-exclusion among online
gamblers: An empirical real-world study. Journal of Gambling Studies, 39, 447–465. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​
.​1​0​0​7​/​s​1​0​8​9​9​-​0​2​2​-​1​0​1​4​9​-​z​​​​

Hopfgartner, N., Auer, M., Helic, D., & Griffiths, M. D. (2024). Using artificial intelligence algorithms to
predict self-reported problem gambling among online casino gamblers from different countries using
account-based player data. International Journal of Mental Health and Addiction, 1–23. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​
/​1​0​.​1​0​0​7​/​s​1​1​4​6​9​-​0​2​4​-​0​1​3​1​2​-​1​​​​
Howard, A., AritraAmex, Xu, D., Vashani, H., inversion, Negin, Dane, S., 2022. American express - default
prediction. ​h​t​t​p​s​:​​/​/​k​a​g​​g​l​e​.​c​o​​m​/​c​o​​m​p​e​t​i​​t​i​o​n​s​​/​a​m​e​x​-​​d​e​f​a​​u​l​t​-​p​r​e​d​i​c​t​i​o​n. Kaggle.
Kairouz, S., Costes, J. M., Murch, W. S., Doray-Demers, P., Carrier, C., & Eroukmanoff, V. (2023). Enabling
new strategies to prevent problematic online gambling: A machine learning approach for identifying
at-risk online gamblers in france. International Gambling Studies, 23, 471–490. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​
0​/​1​4​​4​5​9​7​9​​5​.​2​0​2​2​​.​2​1​6​​4​0​4​2
Ladouceur, R., Shaffer, P., Blaszczynski, A., & Shaffer, H. J. (2016). Responsible gambling: a synthesis of
the empirical evidence. Addiction Research & Theory, 25, 225–235. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​0​/​1​6​​0​6​6​3​5​​9​
.​2​0​1​6​​.​1​2​4​​5​2​9​4
LeCun, Y., Kavukcuoglu, K., Farabet, C., 2010. Convolutional networks and applications in vision, in: Proceedings of 2010 IEEE international symposium on circuits and systems, IEEE. pp. 253–256. DOI:
10.1109/ISCAS.2010.5537907.
Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., Denniston, A. K., Ashrafian, H., Beam, A. L., Chan, A. W.,
Collins, G. S., Deeks, A. D. J., ElZarrad, M. K., Espinoza, C., Esteva, A., Faes, L., Ferrante di Ruffano,
L., Fletcher, J., Golub, R., Harvey, H., Haug, C., … Yau, C. (2020). Reporting guidelines for clinical
trial reports for interventions involving artificial intelligence: the consort-ai extension. The Lancet Digital Health, 2, e537–e548. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​1​6​/​s​2​5​8​9​-​7​5​0​0​(​2​0​)​3​0​2​1​8​-​1
Longjohn, R., Kelly, M., Singh, S., Smyth, P., 2024. Benchmark data repositories for better benchmarking.
Advances in Neural Information Processing Systems 37, 86435–86457. 10.48550/arXiv. 2410.24100.
Louderback, E. R., LaPlante, D. A., Currie, S. R., & Nelson, S. E. (2021). Developing and validating lower
risk online gambling thresholds with actual bettor data from a major internet gambling operator. Psychology of Addictive Behaviors, 35, 921. https://doi.org/10.1037/adb0000628
Marionneau, V., Ristolainen, K., & Roukka, T. (2025). Duty of care, data science, and gambling harm: A
scoping review of risk assessment models. Computers in Human Behavior Reports, 18, Article 100644.
https://doi.org/10.1016/j.chbr.2025.100644
McAuliffe, W. H. B., Louderback, E. R., Edson, T. C., LaPlante, D. A., & Nelson, S. E. (2022). Using “markers of harm” to track risky gambling in two cohorts of online sports bettors. Journal of Gambling Studies,38, 1337–1369. https://doi.org/10.1007/s10899-021-10097-0
Mindway AI, n.d. Award-winning early detection of at risk and problem gambling. ​h​t​t​p​s​:​/​/​m​i​n​d​w​a​y​.​a​i​/​g​a​m​e​
s​c​a​n​n​e​r​/​​​.​ Accessed: 2025–11-25.
Morrow, E. M., Yoo, S. Y., Flavell, S. W., Kim, T. K., Lin, Y., Hill, R. S., Mukaddes, N. M., Balkhy, S., Gascon, G., Hashmi, A., Al-Saad, S., Ware, J., Joseph, R. M., Greenblatt, R., Gleason, D., Ertelt, J. A., Apse,
K. A., Bodell, A., Partlow, J. N., … Walsh, C. A. (2008). Identifying autism loci and genes by tracing
recent shared ancestry. Science, 321, 218–223. https://doi.org/10.1126/science.1157657
Murch, W.S., French, M., Kairouz, S., 2025. Online gamblers’ preferences for performance and fairness in
Artificial Intelligence systems for gambling harm detection. Journal of Gambling Issues ​h​t​t​p​s​:​​/​/​c​d​s​​p​r​e​s​
s​.​​c​a​/​w​​p​-​c​o​n​​t​e​n​t​/​​u​p​l​o​a​d​​s​/​2​0​​2​5​/​0​9​​/​J​G​I​-​​O​c​t​-​2​4​​-​R​E​S​​-​4​8​2​.​​R​3​-​_​M​​u​r​c​h​_​P​​r​o​o​f​​_​F​I​N​A​L​.​p​d​f.
Murch, W. S., Kairouz, S., Dauphinais, S., Picard, E., Costes, J., & French, M. (2023). Using machine learning to retrospectively predict self-reported gambling problems in quebec. Addiction, 118, 1569–1578.
https://doi.org/10.1111/add.16179
Murch, W. S., Kairouz, S., & French, M. (2024). Comparing ‘fair’ machine learning models for detecting
at-risk online gamblers. International Gambling Studies, 25, 84–106. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​0​/​1​4​​4​5​9​7​9​​
5​.​2​0​2​4​​.​2​4​1​​2​0​5​1
Murch, W. S., Kairouz, S., & French, M. (2024). Establishing the temporal stability of machine learning
models that detect online gambling-related harms. Computers in Human Behavior Reports, 14, Article
100427. https://doi.org/10.1016/j.chbr.2024.100427
Nelson, S. E., LaPlante, D. A., Peller, A. J., Schumann, A., LaBrie, R. A., & Shaffer, H. J. (2008). Real limits in the virtual world: Self-limiting behavior of internet gamblers. Journal of Gambling Studies, 24,
463–477. https://doi.org/10.1007/s10899-008-9106-8
Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y., 2011. Reading digits in natural images with
unsupervised feature learning, in: NIPS Workshop on Deep Learning and Unsupervised Feature Learning, p. 7.
Newall, P., & Swanton, T. B. (2024). Beyond ‘single customer view’: Player tracking’s potential role in
understanding and reducing gambling-related harm. Addiction, 119, 1156–1163. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​.​1​1​
1​1​/​a​d​d​.​1​6​4​3​8​​​​

Paprica, P.A., Sutherland, E., Smith, A., Brudno, M., Cartagena, R.G., Crichlow, M., Courtney, B., Loken,
C., McGrail, K.M., Ryan, A., Schull, M.J., Thorogood, A., Virtanen, C., Yang, K., 2020. Essential
requirements for establishing and operating data trusts: Practical guidance based on a working meeting
of fifteen canadian organizations and initiatives. International Journal of Population Data Science 5.
10.23889/ijpds.v5i1.1353.
Parke, A., Harris, A., Parke, J., & Goddard, P. (2015). Understanding within-session loss-chasing: An experimental investigation of the impact of stake size on cognitive control. Journal of Gambling Studies, 32,
721–735. https://doi.org/10.1007/s10899-015-9570-x
Percy, C., 2024. How can we trust what other people say about AI? Presented at the International Association
of Gambling Regulators (IAGR) conference, Rome, October 2024.
Percy, C., Dragicevic, S., Sarkar, S., & d’Avila Garcez, A. (2022). Accountability in AI: From principles to
industry-specific accreditation. AI Communications, 34, 181–196. https://doi.org/10.3233/aic-210080
Percy, C., França, M., Dragičević, S., & d’Avila Garcez, A. (2016). Predicting online gambling self-exclusion: an analysis of the performance of supervised machine learning models. International Gambling
Studies, 16, 193–210. ​h​t​t​p​s​:​​/​/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​8​0​/​1​4​​4​5​9​7​9​​5​.​2​0​1​6​​.​1​1​5​​1​9​1​3
Percy, C., Garcez, A., Dragicevic, S., Sarkar, S., 2020. Lessons learned from problem gambling classification:
Indirect discrimination and algorithmic fairness, in: AI for Social Good - AAAI Fall Symposium 2020.
https://ceur-ws.org/Vol-2884/paper_107.pdf.
Perrot, B., Hardouin, J. B., Thiabaud, E., Saillard, A., Grall-Bronnec, M., & Challet-Bouju, G. (2022). Development and validation of a prediction model for online gambling problems based on players’ account
data. Journal of Behavioral Addictions, 11, 874–889. https://doi.org/10.1556/2006.2022.00063
Philander, K. S. (2013). Identifying high-risk online gamblers: a comparison of data mining procedures.
International Gambling Studies, 14, 53–63. https://doi.org/10.1080/14459795.2013.841721
Samsel, K., Noaeen, M., Seeman, N., Keshavjee, K., Li, L.J., Shakeri, Z., 2025. Multimodal generative ai and
foundation models for behavioural health in online gambling. 10.48550/ARXIV.2503.03752.
Sarkar, S., Weyde, T., Garcez, A.D., Slabaugh, G., Dragicevic, S., Percy, C., 2016. Accuracy and interpretability trade-offs in machine learning applied to safer gambling, in: Proceedings of the NIPS 2016
Workshop on Comprehensible Computational Intelligence (CoCo 2016), Barcelona, Spain. ​h​t​t​p​:​/​​/​c​e​u​
r​​-​w​s​.​o​r​​g​/​V​o​​l​-​1​7​7​​3​/​C​o​C​​o​N​I​P​S​_​​2​0​1​6​​_​p​a​p​e​r​1​0​.​p​d​f.
Schellinck, T., & Schrans, T. (2011). Intelligent design: How to model gambler risk assessment by using
loyalty tracking data. Journal of Gambling Issues, 51. https://doi.org/10.4309/jgi.2011.26.5
Shaffer, H. J., Blaszczynski, A., & Ladouceur, R. (2020). Gambling control and public health: Let’s be honest. International Journal of Mental Health and Addiction, 18, 819–824. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​.​1​0​0​7​/​s​1​1​4​
6​9​-​0​2​0​-​0​0​2​4​0​-​0​​​
Silver, N. (2012). The Signal and the Noise: Why So Many Predictions Fail-but Some Don’t. New York: The
Penguin Press.
Susser, D., Schiff, D. S., Gerke, S., Cabrera, L. Y., Cohen, I. G., Doerr, M., Harrod, J., Kostick-Quenet, K.,
McNealy, J., Meyer, M. N., Price, W. N., & Wagner, J. K. (2024). Synthetic health data: Real ethical
promise and peril. Hastings Center Report, 54, 8–13. https://doi.org/10.1002/hast.4911
Sustainable Interaction, n.d. Collaboration and innovation: The future of safer gambling. ​h​t​t​p​s​:​​/​/​w​w​w​​.​s​u​
s​t​a​​i​n​a​b​​l​e​g​a​m​​b​l​i​n​g​​.​s​e​/​a​r​​t​i​c​l​​e​s​/​c​o​​l​l​a​b​o​​r​a​t​i​o​n​​-​a​n​d​​-​i​n​n​o​​v​a​t​i​o​​n​-​t​h​e​-​​f​u​t​u​​r​e​-​o​f​-​s​a​f​e​r​-​g​a​m​b​l​i​n​g. Accessed:
2025–11-25.
Thiyagalingam, J., Shankar, M., Fox, G., & Hey, T. (2022). Scientific machine learning benchmarks. Nature
Reviews Physics, 4, 413–420. https://doi.org/10.1038/s42254-022-00441-7
Tran, L. T., Wardle, H., Colledge-Frisby, S., Taylor, S., Lynch, M., Rehm, J., Volberg, R., Marionneau, V.,
Saxena, S., Bunn, C., Farrell, M., & Degenhardt, L. (2024). The prevalence of gambling and problematic gambling: a systematic review and meta-analysis. The Lancet Public Health, 9, e594–e613. ​h​t​t​p​s​:​​/​
/​d​o​i​​.​o​r​g​/​1​​0​.​1​0​​1​6​/​s​2​4​6​8​-​2​6​6​7​(​2​4​)​0​0​1​2​6​-​9
UK Gambling Commission, 2021. Consultation: Online games design and reverse withdrawals. ​h​t​t​p​s​:​​/​/​w​w​w​​
.​g​a​m​b​l​​i​n​g​c​​o​m​m​i​s​​s​i​o​n​.​​g​o​v​.​u​k​​/​c​o​n​​s​u​l​t​a​​t​i​o​n​-​​r​e​s​p​o​n​​s​e​/​o​​n​l​i​n​e​​-​g​a​m​e​​s​-​d​e​s​i​​g​n​-​a​​n​d​-​r​e​​v​e​r​s​e​​-​w​i​t​h​d​​r​a​w​a​​l​s​/​o​g​​
d​r​w​-​p​​r​o​h​i​b​i​​t​i​o​n​​-​o​f​-​r​​e​v​e​r​s​​e​-​w​i​t​h​​d​r​a​w​​a​l​s​-​f​o​r​-​a​l​l​-​r​e​m​o​t​e​-​o​p​e​r​a​t​o​r​s. Accessed: 2025–11-25.
UK Gambling Commission, 2023. Statistics and research release: Risk algorithms data. ​h​t​t​p​s​:​​/​/​w​w​w​​.​g​a​m​b​l​​i​n​
g​c​​o​m​m​i​s​​s​i​o​n​.​​g​o​v​.​u​k​​/​s​t​a​​t​i​s​t​i​​c​s​-​a​n​​d​-​r​e​s​e​​a​r​c​h​​/​p​u​b​l​​i​c​a​t​i​​o​n​/​r​i​s​​k​-​a​l​​g​o​r​i​t​h​m​s​-​d​a​t​a. 26 April 2023.
Vanderstichele, H., De Meyer, G., Shapiro, F., Engelborghs, S., De Deyn, P.P., Shaw, L.M., Trojanowski,
J.Q., 2008. Alzheimer’s disease biomarkers: from concept to clinical utility. Biomarkers for early diagnosis of Alzheimer’s Disease, 81–122.
Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., & Ni, B. (2023). Medmnist v2 - a large-scale
lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10. ​h​t​t​p​s​:​/​/​d​o​i​.​
o​r​g​/​1​0​.​1​0​3​8​/​s​4​1​5​9​7​-​0​2​2​-​0​1​7​2​1​-​8​​​​

Zendle, D., & Newall, P. (2024). The relationship between gambling behaviour and gambling-related harm:
A data fusion approach using open banking data. Addiction, 119, 1826–1835. ​h​t​t​p​s​:​/​/​d​o​i​.​o​r​g​/​1​0​.​1​1​1​1​/​
a​d​d​.​1​6​5​7​1​​​​
Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a
publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.