Responsible gambling AI: how operators try to spot harm

A plain guide to the models that scan betting accounts for signs of harm, what the Gambling Commission expects of its licensees, what the research shows & where the controversy lies.

What ‘responsible gambling AI’ means

“Responsible gambling AI” is a loose label for software that reads a customer’s betting account data & tries to flag signs of gambling harm, so that the operator can step in. Online gambling suits this approach because, as researchers Michael Auer & Mark Griffiths note in a Journal of Gambling Studies paper, every transaction is assigned to one account & recorded.

The terms are often used interchangeably, but Auer & Griffiths draw a distinction. Machine learning refers to “a group of advanced statistical methods”, while AI “can be regarded as the outcome of an advanced algorithm”.

These algorithms need a training dataset: a set of past examples from which the model learns the patterns of a chosen group, such as people who report problem gambling. Once trained, a model can score new accounts against those patterns.

Regulators have noticed the trend. In a keynote to the International Association of Gaming Regulators on 20 October 2025, Gambling Commission chief executive Andrew Rhodes said artificial intelligence “is starting to make itself felt”. He said that in the 12 months or so before that speech, operators in Great Britain had increasingly used generative AI to try to improve the consistency of customer interactions.

What the Gambling Commission requires

In Great Britain, the Gambling Commission’s customer interaction guidance for remote gambling licensees is formal guidance under social responsibility code provision 3.4.3. The page was last updated on 23 August 2023, following a consultation on remote customer interaction.

Its “Identify” section sets out several requirements. Requirement 4 states: “There are effective systems and processes to monitor all customer activity and identify harm.” Requirement 5 sets “seven relevant categories of indicators” that licensees must use to help identify gambling-related harm.

Requirement 7 adds a sense of urgency: “The greater the harm identified the more important it is to take swift action.” Requirement 3 covers support for customers in a vulnerable situation, & Requirement 6 says customers must get the same level of protection where a licensee contracts with third parties.

The requirement titles summarised here do not name a particular technology. They describe the outcome: effective systems & processes to monitor all customer activity & identify harm.

How the models are built

Auer & Griffiths, in their paper published on 19 July 2022, describe a typical approach. They were given raw data on 1,287 players from a European online casino who answered questions on the Problem Gambling Severity Index (PGSI), a problem gambling screen, between September 2021 & February 2022.

The players’ PGSI answers served as the label, meaning the outcome the model tries to predict. The inputs, or features, were drawn from account behaviour, with the authors giving wagering, depositing & gambling frequency as examples.

Self-reported screens are not the only possible label. The paper notes that several earlier studies used AI methods to predict voluntary self-exclusion, where customers bar themselves from gambling, & that one study predicted voluntary limit setting among Norwegian online players.

The data showed distinct patterns. According to the authors, problem gamblers lost more money per gambling day & per session, deposited more often per session & tended to deplete their accounts more frequently. A subgroup identified as being at greater harm showed even higher values on these measures.

How accurate are they?

Auer & Griffiths trained two types of algorithm, random forest & gradient boost machine. They report that the random forest model predicted self-reported problem gambling better than gradient boost.

The authors conclude that “self-reported problem gambling can be predicted by AI algorithms with high accuracy based on player tracking data”. The source text available for this guide does not carry the specific accuracy scores, so they are not reported here.

Two limits follow from the study design as described. The model predicts what players said about themselves on a questionnaire, not a clinical diagnosis, & the data came from a single European online casino.

Not everyone in the industry accepts that such models help. As set out below, DraftKings’ chief responsible gaming officer has said the company’s evidence showed risk-modelling technology was not helpful.

In practice: player-facing tools

Some tools are aimed at players directly rather than at back-office risk teams. Yogonet reported that DraftKings was adding custom cool-offs, letting customers choose a break from gambling of any length between three & 364 days.

DraftKings also launched Gamalyze American Football with Mindway AI, described as “a gamified responsible engagement experience” for American football fans. According to the report, this built on DraftKings’ integration of Gamalyze Casino into its My Budget & Controls area in January 2026 & a soccer-themed Gamalyze Football during the 2026 FIFA World Cup.

The company listed other tools, including My Budget Builder, My Stat Sheet & player-set limits. According to Yogonet, DraftKings was also launching a PGA TOUR sweepstakes, running “from September 16 through September 29”, that gives “eligible customers the opportunity to win an all-expenses-paid trip” by engaging with features such as My Budget Builder & My Stat Sheet.

“At DraftKings, responsible engagement is embedded in how we operate and is essential to building trust with our customers and our long-term sustainability,” said Lori Kalani, the company’s chief responsible gaming officer. The report does not describe how Gamalyze works internally.

When the same data sells

The account data that can reveal harm can also show who responds to offers. Gizmodo, summarising a New York Times investigation based on interviews with more than 40 former DraftKings employees, reported that in 2023 DraftKings built a model scoring users on “elasticity”, a measure of how they respond to free bets & promotions.

One former employee, Jayden Butts, told the Times the model asked: “Is this person going to give us more than we’re giving them?” If the answer was yes, he said, the company would “open the floodgates” with promotions.

Former employees also described a separate model designed to predict when users were heading towards a crisis & needed an intervention. According to the Times, a planned presentation of that model in early 2025 was cancelled, & two other attempts to build similar algorithms were shelved.

DraftKings told the Times it “rejects any implication that its marketing practices are unfair or improperly targets customers”, saying promotions go to customers who show “sustained, engaged use”, not to customers based on their losses. Kalani said the company monitors for “potentially risky behaviors” but that evidence had shown risk-modelling technology was not helpful.

According to Covers, the Massachusetts Gaming Commission said it would review how all its licensed sportsbooks use AI, not only DraftKings. Covers notes the commission has not determined that DraftKings committed any wrongdoing.

The review will draw on an AI task force led by executive director Dean Serpa, formed after a 2025 study by the University of Nevada Las Vegas’ International Gaming Institute found a “governance gap” between technology & existing practices & regulation. Commissioner Paul Brodeur said the challenge lies in “determining the potential for AI to potentially exploit customers”.

Why monitoring matters: enforcement

Enforcement cases show what regulators treat as a failure to spot harm. On 7 October 2026, the Gambling Commission announced that three Rank Group operators, Grosvenor Casinos Limited, Grosvenor Casinos (GC) Limited & Gaming Group Limited, would pay £5 million for anti-money laundering & social responsibility failures.

The social responsibility failures included no safer gambling interactions with a customer during a period in which they lost £50,000. There was no record of any interaction with a customer who won about £260,000 in a short period & then lost around £250,000 in 12 days.

A customer returning after self-exclusion had no safer gambling interaction until they had lost £25,000. The operators, which run 51 casinos across Great Britain, will also undergo a third-party audit, & the full £5 million goes to the Government’s Consolidated Fund.

The Commission described the operators as land-based casino operators. Sue Young, the Commission’s executive director of operations, said the risks “are equally alive in the land-based sector”.

What good practice might involve

The sources do not set out a single standard for responsible gambling AI, but they point to recurring questions. Rhodes described how, over several years, operators moved from having no policies for meaningful interactions to having policies they did not always follow, then to interactions that happened but were often inconsistent in quality.

Rhodes also warned about hyper-personalisation, saying the industry needs to balance “serving meaningful content” against “people are starting to feel an increased intensity that nobody necessarily intended”. The DraftKings allegations show why that balance matters when harm models & marketing models draw on the same data.

Auer & Griffiths argue that good tracking tools should support informed player choice & also help operators understand players’ behaviour. The Commission’s Requirement 6 is relevant where monitoring relies on outside suppliers, because customers must receive the same protection.

Measuring whether any intervention works is hard. Rhodes noted that changes usually arrive as packages, making it difficult to disaggregate the effect of each one, a point that applies to judging tools as much as rules.

Questions

Does the Gambling Commission require its remote licensees to use AI?

The customer interaction guidance for remote licensees requires effective systems & processes to monitor all customer activity & identify harm, using seven categories of indicators. The guidance summarised here does not specify a technology.

What data do harm-detection models use?

In the Auer & Griffiths study, models used account data such as wagering, depositing & gambling frequency, trained against players’ answers on the Problem Gambling Severity Index. Other studies have used voluntary self-exclusion as the outcome to predict.

What is DraftKings accused of?

A New York Times investigation, as summarised by Gizmodo, reported that DraftKings built an ‘elasticity’ model to guide promotions while internal crisis-prediction efforts were shelved. DraftKings rejects any implication that its marketing practices are unfair or improperly target customers.

What did Rank’s casinos fail to do?

According to the Gambling Commission on 7 October 2026, the operators carried out no safer gambling interactions with one customer who lost £50,000 & did not interact with a returning self-excluder until they had lost £25,000. They agreed to pay £5 million.

Sources

• gamblingcommission.gov.uk
• link.springer.com
• yogonet.com
• gizmodo.com
• covers.com
• gamblingcommission.gov.uk
• gamblingcommission.gov.uk

Sources checked 11 October 2026. Information only. 18+.