AI in Online Gambling: Opportunities, Risks, and Ethical Dilemmas

Perhaps the best thing AI can do to improve the gambling industry is to identify problem gamblers based on their betting history or customer support logs and help them receive the help they require.

In one Swedish study, for instance, (Auer et al, 2020) researchers were given access to a dataset of 7,134 gamblers whose gambling behavior was tracked by a behavioral feedback system that sent them personalized messages that encouraged responsible gaming practices based on rules and machine learning algorithms.

The study found that these messages based on behavior such as high losses, increased deposits, and greater gambling duration had a significant impact on behavior in the immediate term.

 In total, 65% of players reduced their gambling activities on the day they read a message and 60% reduced their betting seven days after the message. The effect was slightly reduced for players the tracking tool classified as ‘high-risk’, but still potent.

Another means for responsible gambling intervention with AI assistance is examining customer service interactions to find indicators of problem gambling. 

Using a Linguistic Inquiry and Word Count tool to evaluate 1008 emails researchers (Haefeli et al 2014) found that automated text analysis could be predictive of gambling issues leading to self-exclusion.

This LIWC evaluated interactions on a number of scales relating to word usage, and found a few that were related to problem gambling behavior. For instance, words that showed up on its anger and time scales were positively predictive of future self-exclusion, while causation was negatively predictive.

Considering the massive volume major betting companies have with customers — and the time-consuming nature through emails and chat logs manually — automatic text analysis could be a helpful tool for problem gambling interventions. 

That said, at the time of the study the LIWC was still found to be “inferior to a human assessors” and it will likely take both a combination of AI analysis and human intervention to achieve the best outcomes.

Although customer service advancements may not be as consequential to bettors’ lives as preventing problem gambling, they can be a win-win for consumers and businesses. 

First and foremost it is a cost saving tool for online betting platforms that reduce the cost of a customer interaction by more than 80% (Muro and Andes, 2015). It also makes it more feasible for companies to offer a wider breadth of customer service support, regardless of their field.

A recent case study of a Brazilian commercial bank (De Andrade and Cleonir Tumelero, 2022) included 433 minutes of interviews with bank employees, many of whom were effusive in their praise of the AI customer service.

We can change a service that would take, I suppose, 20 minutes at an agency, into seconds. Hence, the efficiency is very high; it is not possible to compare a well-done service with the chatbot, to a slower service.

That study pointed to an increase in efficiency that wasn’t only welcome, but necessary. In 2021, there were 400 million interactions, an increase of about 2,350% when compared to 2019 (17 million interactions). More than 15.9 million customers used the virtual assistant in 2021, an increase around 1,590%, compared to 2019 (one million customers). 

It would be easy to assume an enormous surge like that would include a drop in satisfaction, but just 7.9% of these interactions ended up being transferred to a human assistant. We don’t have the same publicly available data for any major sportsbooks, but it would not be surprising if the results were similar.

Dealing with AI chatbots isn’t always associated with the best customer experience, but there is evidence to suggest that there are instances where it is actually preferable to dealing with a human. When dealing with low-complexity tasks customers rated AI chatbots’ problem solving as superior to humans and showed a preference for using them (Xu et al 2020).

Another area where consumers and businesses can both win with AI usage is fraud reduction. 

One thing that artificial intelligence does particularly well is identifying patterns and anomalies, which is at the heart of identifying any kind of financial fraud. 

For example, AI has already demonstrated an impressive aptitude for detecting money laundering, which has often been a concern in the gambling industry.

A recent study (Lebanca et al 2022) tested an active learning framework known as ‘Amaretto’ to see if it could outperform ‘state-of-the-art solutions’ by ‘reducing money laundering risk and improving detection performance’ with promising results.

The study came to the conclusion that Amaretto improved detection rate by 50% and reduced cost by 20%, but perhaps the most impressive finding was the system’s continual improvement. Amaretto doubled its Precision – defined as the proportion of anomalous high level vectors correctly classified as anomalous — in approximately 10 days, reaching a score of 0.78.

• AI success rates for identifying financial anomalies can sometimes be even stronger in gambling-specific cases. A 2022 study that investigated the effectiveness of machine learning techniques to flag fraudulent iGaming behavior (Farrugia, D., Zerafa, C., Cini, T. et al. 2022) reached an average precision rate of 84.2%.

Is AI Encouraging Problem Gambling?

Many of the drawbacks of AI in the gambling industry are more theoretical than concrete. There isn’t necessarily evidence that this technology is being used for nefarious purposes, but the potential is undoubtedly there.

The studies listed above about AI identifying possible problem gamblers for productive responsible gambling interventions based on betting patterns and gambling patterns could easily be weaponized. It is just as easy for an AI to send a message encouraging further betting via personalized advertisements or promotions as it is for it to deliver one that suggests a cool-off period if it is programmed in an unethical manner.

Not only is there no shortage of ways to identify vulnerable individuals using AI, it can also be programmed with knowledge of what motivates people to bet — like the strong correlation (r = .35) between financial motives and problem gambling (Tabri, N., Xuereb, S., Cringle, N., & Clark, L. 2021).

• Youth gambling is of particular concern when it comes to AI enticement. It is difficult to get reliable statistics on it because it is often illegal and most data relies on self reporting. A recent Spanish study (Rial et al, 2022) found that 23.5% of children aged 12-17 had gambled in their lifetime with 1.9% presenting signs of problem gambling. A similar Scottish study found that 26% of children 11-17 had spent their own money on gambling in the previous 12 months and 0.7% were already identified as problem gamblers.

Much of the potential damage in this realm is theoretical, but bad-faith actors using AI have a dangerous tool at their disposal.

One common concern with AI is its capability to compromise user privacy. This is an issue most associated with AI assistants that deal with a massive amount of a single consumer’s data, but that doesn’t mean it isn’t a concern in other contexts.

There are emerging concerns about how improving AI chatbots might make users more likely to disclose sensitive information (Kronemann et al 2022), but a more tangible threat is likely AI-related data breaches.

This is a problem across a variety of industries, but could be particularly harmful in the gambling space considering the financial information betting platforms can have about their customers. A Hidden Layer study on AI security threats found that 77% of companies had faced an AI security breach of some kind, and the company’s CEO laid out his opinion of the threat in no uncertain terms.

Artificial intelligence is, by a wide margin, the most vulnerable technology ever to be deployed in production systems. It’s vulnerable at a code level, during training and development, post-deployment, over networks, via generative outputs, and more.” – Chris Sestito, founder and CEO of HiddenLayer 

This worry is a broad one, but for customers who may be trusting betting platforms with large sums of their money it is particularly disconcerting. 

Casinos have already been the targets of significant cyber attacks recently, and it’s possible that AI-related vulnerabilities could continue to make them targets.

AI Ethical Concerns

This research has touched on some of the ethical concerns related to AI use in gambling on the operator side, but one way to conceptualize the difficulties that arise as a result of this new technology is to examine ethics guidelines that currently exist.

A fascinating meta-analysis published in Nature Machine Intelligence (Jobin et al, 2019) did just that. It dove into 84 ethical guidelines for AI usage from around the world in search of common themes providing insight about universal concerns surrounding artificial intelligence.

Below is a chart with the ethical principles that appeared most often in these guidelines:

Some of these concerns are fairly general — such as ‘Justice and Fairness’ or ‘Responsibility’ — but others are more specific and applicable to the betting industry. 

One significant takeaway from this meta-analysis is that however iGaming companies use artificial intelligence there is a strong demand for transparency on that account. Maleficence and breaches of trust are much more likely to occur in the dark than out in the open. 

Keeping the principles above in mind is one way for betting companies to ensure their use of AI remains ethical.

Much of this research has focused on how AI changes the way betting platforms interact with their customers, but it also has an effect on bettor behavior.

Sports betting advice is a use case for ChatGPT that bettors are clearly interested in, and there are a variety of freely-available GPTs built for that purpose. Putting the power of AI towards wagers may seem like a compelling concept to some bettors, but most of these options are highly limited. 

For example, asking ‘Betting Pros AI’ to provide picks on a particular game with explanations results in vagaries like this:

For reference, at the time of these picks Leo Jimenez had never hit an MLB home run and this pick seemed to be a reflection of the odds alone. Alejandro Kirk had strong recent form, but the explanation did not reflect that.

Another similar GPT named ‘Sports Betting Master’ had an even harder time producing picks, not even specifying odds and instead spitting out general concept that weren’t accurate:

A couple of issues included:

• The Orioles had won 7 of their 11 at the time of the recommendation, hardly ‘shaky form’ while the Blue Jays ‘hot streak’ was non-existent as they’d won 3 of their last 7.
• Both teams were said to have ‘decent bullpens’ as a rationale for an under bet even though the AI had just said the Orioles’ relievers were fatigued — and Toronto had the second-worst bullpen in MLB by ERA at the time.
• Kevin Gausman wasn’t starting that day.

These are very basic examples, and there are plenty of claims that ChatGPT works for betting on platforms like YouTube, but they are largely unverifiable. There are also examples of bettors struggling to get actionable advice or losing a large percentage of their starting bankroll.

Because ChatGPT is now able to use current data, so it has theoretical betting applications — and there are those who claim to have been able to use it in this way effectively — but as of writing it is likely better used as a tool for line shopping, and perhaps finding arbitrage betting opportunities, than picking bets with strong value.

For now, the greater AI betting opportunity lies with advanced machine learning models that are not necessarily available to the average better. For years high-information bettors have been creating models to direct their betting activity and AI technology has only granted more horsepower to this pursuit.

It’s not uncommon for advanced statistical models to yield impressive results when it comes to predicting sports results. Examples with basketball alone include (compiled by Hubáček et al 2019):

• Loeffelholz, Bednar, and Bauer (2009) achieving an accuracy of over 74% using neural network models, however their dataset consisted of only 620 games
• Ivanković, Racković, Markoski, Radosav, and Ivković (2010) using advanced neural networks to predict outcomes of basketball games in the League of Serbia in seasons 2005/06–2009/10 with 81% accuracy.
• Puranmalka (2013) using play-by-play data and a support vector machine to achieve accuracy over 71% for 10 NBA seasons from 2003/04 to 2012/13.

Accuracy is not the only important metric when it comes to statistical modeling effectiveness in the sports betting space. 

While bettors hope to win a high percentage of their wagers, final return on investment is what matters the most. On that count a recent study (Walsh et al, 2024) found that a calibration-based model that focused on the confidence of its predictions rather than just whether they were correct yielded a +32.45% ROI while that number dropped to +5.56% for an accuracy-based model. This finding could go a long way towards improving future attempts to create profitable sports-betting models.

As AI gets more advanced its possible results like this will get even stronger, but it remains to be seen how accessible profitable AI-driven betting models will be for the average bettors and how sportsbooks may look to counter their impact.

With AI tools relatively new and unfamiliar to many bettors, it’s difficult to make any broad judgments about consumer trust in them.

In different contexts there is some evidence to suggest that trust is fairly high. For instance, one study (Kim et al, 2021) gave 157 participants AI recommendations for ‘low-quality’ and ‘high-quality’ books — ie. books that were top-20 on bestseller lists or not — and had them self-report their purchase intention and trust in the AI recommendation.

All of the numbers below are rated on a scale of 1-to-7 and the difference between the ‘imprecise’ and ‘precise’ conditions is that the AI recommendations came with a confidence interval that was either a round number (80% - imprecise) or one with multiple decimal places (79.865%). 

Here are the ratings the participants gave:

Buying a book is a different proposition than making a bet, but it still demands a financial commitment from the consumer and if the 1-to-7 scale is converted to percentages the trust and purchase intention never dipped below 59.4% in any condition — and climbed as high as 74.0%.

Consumer perception of AI is sure to evolve as the technology does, but there appears to be a relatively high floor for the acceptance of artificial intelligence recommendations, even when they involve a financial component.

Artificial intelligence is moving too quickly for there to be definitive judgments made on whether it is a net-positive or negative for the betting industry.

Its capacity to assist in responsible gambling interventions is immensely promising, but as is the possibility for unscrupulous applications that target vulnerable bettors.

At that moment its functionality as a tool for the average bettor is limited, but there are those for whom it may already be indispensable. 

If a time comes when accessible, free, AI tools are consistently capable of finding value against sportsbooks it could be immensely disruptive to the industry — but we are not on the precipice of that yet, and online betting platforms will undoubtedly be using the most sophisticated technology available to prevent it from happening.

In most discussions about AI, concerns lie with how it will be used rather than what it’s capable of, and the iGaming industry is no exception. It could easily help promote responsible gambling while making betting platforms more efficient, but it would be foolish to ignore the possible drawbacks associated with it.