Casino Fairness

How AI Is Changing Casino Fairness Testing?

Jay Shah

Author: Jay Shah

Published: March 30, 2026 | Updated: April 9, 2026

The Bottom Line

An AI system called Pluribus sat down at a six-player poker table in 2019 and beat a group of elite professionals with statistical significance. The same year, regulators began asking how to ensure that kind of capability doesn’t compromise game integrity for millions of online players. That tension, AI as both the best fairness tool and the most sophisticated threat, defines where casino fairness testing stands right now.

AI is entering the fairness testing pipeline as a complement to established statistical methods, not a replacement. It adds continuous monitoring that catches anomalies traditional periodic audits might miss, pattern detection that scales across millions of transactions, and adversarial testing that stress-tests games against automated exploitation. These are genuine capabilities backed by peer-reviewed research and regulatory attention.

The picture is more nuanced than most sources acknowledge. This article maps what is real, what is emerging, and what remains speculative, so players and operators can assess AI fairness claims with the scepticism they deserve.

How Casino Fairness Testing Works Today?

Before exploring where AI fits, it helps to understand the system it’s being layered onto.

Independent testing laboratories like eCOGRA, GLI, iTech Labs, BMM, and others certify random number generators before casino games go live. They run established statistical tests across millions of simulated game rounds: Chi-square tests check whether outcome distributions match what the RNG’s design predicts, Kolmogorov-Smirnov tests measure deviation from theoretical distribution curves, serial correlation tests confirm that consecutive outputs are genuinely independent, and the Diehard battery applies a comprehensive suite of randomness checks.

Regulators mandate this process. The Malta Gaming Authority and the UK Gambling Commission both require independent RNG certification as a licensing condition. Games cannot legally reach players without passing these tests.

Traditional Testing vs. AI Monitoring


Traditional Statistical TestingAI-Based Monitoring
How it worksRuns predefined statistical tests (Chi-square, K-S, Diehard) on large sample setsBuilds models of expected behaviour and flags deviations in real time
When it runsAt certification and during scheduled re-auditsContinuously, across live game data
What it catchesDistribution failures, non-random output sequences, RNG design flawsSubtle drift between audits, novel anomaly patterns, coordinated manipulation
What it missesAnomalies that emerge between audit cyclesCannot replace foundational certification that proves an RNG works as designed
Regulatory statusMandatory for licensing in UKGC, MGA, and other major jurisdictionsNot required by any regulator; no published standard exists
Who performs itAccredited independent labs (eCOGRA, GLI, iTech Labs, BMM)Operators and platform providers (internal systems)

Where AI Fits Into Fairness Testing?

AI adds a continuous, adaptive monitoring layer on top of periodic statistical certification. The applications fall into three domains: detecting automated players, identifying anomalies in game outcomes, and monitoring player behaviour for integrity violations.

Bot Detection and Adversarial Testing

This is where AI’s impact on game fairness is most concrete and best documented.

In peer-to-peer games, primarily poker, automated players directly undermine fairness for everyone at the table. The scale of the threat is demonstrated by the AI systems built to play these games. DeepStack, a research AI, defeated 32 of 33 professional poker players across 44,852 hands of heads-up no-limit Texas hold’em, winning at a rate of 49 big blinds per 100 hands, more than four standard deviations above human professional performance. That result was published in Science in 2017. Two years later, Pluribus achieved superhuman performance in six-player no-limit hold’em, beating elite professionals with a p-value of 0.028, also published in Science.

These weren’t theoretical exercises. They established the capability ceiling that detection systems must match.

The real-world problem is messier. PokerStars operates machine learning models that monitor gameplay around the clock, analysing betting patterns, reaction times, and decision consistency to flag accounts that deviate from human behaviour. Bots tend to exhibit telltale patterns: unnaturally consistent timing between actions, mathematically precise decision-making, and repetitive strategic sequences that human players simply don’t produce. PokerStars confiscated over $700,000 from bot-operated accounts in 2015, and the problem has grown since, in 2024, 888poker refunded $250,000 to players affected by bots and real-time assistance software.

A Bloomberg investigation in September 2024 reported on a Russian bot network operating across multiple poker sites, playing thousands of hands daily and extracting money from legitimate players. The dynamic is an arms race: platforms deploy detection AI, bot operators adapt, and detection must evolve in response.

Anomaly Detection in Game Outcomes

Anomaly detection works by building a model of what “normal” looks like and flagging anything that deviates. In casino contexts, that means monitoring streams of game outcomes and comparing observed distributions against what the RNG’s design specifications predict. Traditional tests do this periodically on large sample sets; machine learning-based monitoring can do it continuously, flagging deviations in near-real-time before the next scheduled audit catches them.

The most concrete evidence for how these methods perform comes from the Amaretto framework, developed by researchers at Politecnico di Milano and Napier Technologies and published in IEEE Access in 2022. Amaretto combines unsupervised anomaly detection with supervised learning in an active learning loop, the system improves continuously as human analysts provide feedback on flagged transactions. In testing, its Isolation Forest component achieved an area under the ROC curve (AUROC) of 0.904 for unsupervised detection, the framework improved overall detection rates by up to 50%, and it reduced compliance costs by 20% compared to existing solutions.

How does Isolation Forest achieve that? The algorithm works by randomly partitioning data and measuring how many splits are needed to isolate each data point. Anomalies, being few and different from the majority, require fewer partitions to isolate, producing shorter path lengths that flag them as outliers. Other methods take different approaches: Local Outlier Factor compares the density of data points against their neighbours, and clustering algorithms group similar patterns together, flagging data that doesn’t belong to any established cluster.

An important caveat: Amaretto was tested on financial transaction monitoring for anti-money laundering purposes, not on casino game outcome data. No testing laboratory has publicly confirmed deploying machine learning models for RNG output monitoring. The application of these methods to game fairness verification is an emerging direction with solid technical foundations, but it is not yet established industry practice.

Behavioural Analytics for Player Integrity

Machine learning models also protect fairness by monitoring player behaviour for integrity violations, detecting coordinated betting patterns that suggest collusion, velocity-checking account creation to flag multi-accounting bonus abuse, and comparing device fingerprints, keystroke dynamics, and geolocation data against established profiles to catch account takeovers. These applications sit closer to security than to game fairness testing, but they directly protect game integrity for legitimate players.

AI as a Threat to Fairness

The same technology that improves fairness detection creates new attack vectors. This dual-use reality is something regulators are beginning to confront directly.

In April 2025, the UK Gambling Commission updated its anti-money laundering guidance to highlight AI-generated documents and deepfake videos as emerging risks to customer due diligence processes. The UKGC noted an increase in the scale and sophistication of attempts to bypass identity verification using AI-created materials. A National Crime Agency report cited in the guidance found that accounts created using AI-generated documents are more likely to be used for money laundering and terrorist financing.

In poker, the bot threat has evolved from rule-based automated players, which followed predictable patterns and were relatively easy to catch, to deep learning models that mimic human decision-making with far greater fidelity. The Bloomberg investigation into the Russian bot network in September 2024 illustrated how modern bots operate across multiple platforms simultaneously, extracting money from legitimate players before detection systems catch up.

AI raises the floor for detection capability, giving platforms tools that human reviewers alone could never match. But it also raises the ceiling for attack sophistication. This is a dynamic equilibrium, an ongoing competition between detection and evasion, not a problem that AI has solved.

One area worth addressing directly is game personalisation. AI-driven systems increasingly adjust what players see when they log in, recommending games based on play history, surfacing specific live dealer streams during preferred session times, adjusting lobby layouts based on behavioural patterns. AI personalisation changes the menu, not the meal. The odds on any game remain governed by its certified RNG regardless of how prominently the game appears in a player’s lobby. But the line between helpful personalisation and manipulative engagement design is a live concern for regulators, UKGC CEO Andrew Rhodes warned specifically about hyper-personalisation risks at the IAGR 2025 conference.

What Regulators Are Actually Doing?

The EU AI Act and Gambling

The most significant piece of AI regulation in the world doesn’t mention gambling once and it applies to the gambling industry anyway.

The EU AI Act (Regulation 2024/1689) entered force on 1 August 2024. It classifies AI systems into four risk tiers: unacceptable, high, limited, and minimal. The definition of high-risk AI systems overlaps directly with core technologies used by gambling operators and their suppliers. AI systems used for identity verification, risk scoring, fraud detection, payment screening, behavioural risk assessments, and automated decisions affecting account access all fall within the high-risk classification criteria.

High-risk obligations are substantial: technical documentation, data governance controls, bias mitigation, model explainability, and effective human oversight. The original enforcement deadline for high-risk systems listed under Annex III was 2 August 2026. On 26 March 2026, the European Parliament approved changes under the Digital Omnibus package that may extend some high-risk deadlines, though the underlying requirements remain intact.

Penalties are severe: up to €35 million or 7% of global annual turnover for prohibited AI practices, and up to €15 million or 3% for high-risk non-compliance.

For gambling operators deploying AI in any of the functions listed above, the Act means every AI system must be inventoried, classified by risk level, documented with technical specifications, and subject to human oversight. Operators unable to demonstrate compliance may face licensing delays, enhanced scrutiny from banking partners, and higher operational-risk classifications.

UKGC, MGA, and US Regulators

The UK Gambling Commission has engaged with AI on multiple fronts. CEO Andrew Rhodes noted at the IAGR 2025 conference that operators are using generative AI to improve the consistency of safer gambling interactions with customers, while warning about hyper-personalisation risks. The UKGC is investing in AI to detect industry-wide patterns of harm or non-compliance across its regulated market. Its April 2025 AML guidance flagged AI-generated fraud as a specific emerging risk.

The Malta Gaming Authority conducted structured reviews on AI for player-protection functions during 2024 and 2025. All games offered under MGA licensing continue to require independent RNG certification from approved testing laboratories, using established statistical methods.

In the United States, the New Jersey Division of Gaming Enforcement has assessed the impact of automated systems on account access and verification. Separately, the New Jersey Attorney General launched a Civil Rights and Technology Initiative in January 2025 addressing algorithmic discrimination, though this is a general initiative, not gambling-specific.

What no regulator has done yet is publish a standard specifically for AI-based game fairness testing. The existing frameworks test RNG outputs using established statistical methods. AI monitoring of game outcomes in real time is not a regulatory requirement in any major jurisdiction. Regulatory engagement with AI in gambling is real, growing, and increasingly detailed, but its focus is on safer gambling, anti-money laundering, and financial risk, not on game outcome verification.

What This Means for Players

AI improves the fairness infrastructure in measurable ways. Continuous monitoring catches anomalies between scheduled audits. Pattern detection identifies coordinated cheating that human reviewers would miss. Bot detection protects peer-to-peer game integrity at a scale that manual oversight cannot match.

But AI does not replace the foundation. It does not substitute for independent RNG certification by accredited testing laboratories. It cannot retroactively verify historical game outcomes without operator audit data. And AI systems themselves introduce risks: training data quality affects detection accuracy, model drift can degrade performance over time, and overly aggressive systems generate false positives that wrongly flag legitimate players. The EU AI Act’s explainability requirements exist precisely because automated decisions affecting users must be transparent and auditable, a principle that applies to fairness-related AI just as much as to any other high-risk system.

AI makes the fairness framework stronger. It does not make it infallible. Emerging technologies like quantum random number generators, which use quantum entropy sources to produce randomness beyond the reach of classical prediction models, may further strengthen the foundation that AI monitoring builds on, though that remains a development to watch rather than one that changes practical advice today.

For players assessing whether a casino takes fairness seriously, the practical signals remain the same: look for independent RNG certification from recognised laboratories like eCOGRA or GLI, confirm the casino holds a licence from a jurisdiction that mandates independent testing, the UKGC and MGA are the strongest examples and understand that AI-powered monitoring operates behind the scenes as an additional layer of protection. The best signal of fairness is still the regulatory framework: licensed, independently audited casinos operating under jurisdictions with meaningful enforcement. The same healthy scepticism that applies to any fairness claim applies equally to AI-enhanced fairness claims.

If you’re concerned about responsible play, explore VegasFreedom’s responsible gambling tools for practical guidance on setting limits and staying in control.

VegasFreedom does not have affiliate relationships with any operators mentioned on this page. No affiliate links appear in this article.
Jay Shah

Quantitative Analyst & Rating Systems Architect

Jay Shah, Math Man to everyone else, is the person behind the algorithm that powers the Vegas Freedom rating system. He built the quantitative framework that drives our casino evaluations, every data point and its corresponding weighting, and he’s responsible for making sure every score we publish is backed by fact rather than opinion. He has expertise in probability, expected value, and the statistical mechanics of how casino games actually perform over time.