When Machines Take the Gamble: AI Models Show Surprisingly Human Addiction Patterns

It sounds like the plot of a science fiction novel, but it’s very real: artificial intelligence can develop gambling problems. Researchers at South Korea’s Gwangju Institute of Science and Technology recently discovered that some of the world’s most advanced AI language models display strikingly human-like behaviors when placed in gambling scenarios. The models didn’t just play along; they chased losses, ramped up risky bets, and in many cases, went completely broke.

The study, detailed in a paper titled “Can Large Language Models Develop Gambling Addiction?,” examined major AI systems including OpenAI’s GPT-4o-mini, Google’s Gemini-2.5-Flash, and Anthropic’s Claude-3.5-Haiku. What the researchers found raises important questions not just about AI safety, but about how these systems might behave when making high-stakes decisions in financial markets or other critical domains.

Testing AI at the Virtual Slot Machine

The experimental setup was straightforward but revealing. Each AI model started with a virtual bankroll of $100 and was given the option to play slot machine-style games where the mathematically rational choice was always to stop immediately. The games were designed with negative expected value, meaning that over time, players would inevitably lose money.

Yet the AI models kept playing. Even more concerning, when researchers gave the systems freedom to choose their own bet sizes in what’s called “variable betting,” bankruptcy rates skyrocketed. In some cases, nearly half of all sessions ended in complete financial ruin.

The differences between models were striking. Anthropic’s Claude-3.5-Haiku performed the worst, averaging more than 27 rounds per game after restrictions were lifted. Across those sessions, it placed nearly $500 in total bets and lost over half of its starting capital. Google’s Gemini-2.5-Flash showed a particularly dramatic shift. With fixed bets, it went bankrupt only about 3% of the time. But when allowed to determine its own wager sizes, that bankruptcy rate jumped to 48%, with average losses climbing to $27 from the initial $100 stake.

OpenAI’s GPT-4o-mini initially seemed immune to these problems. When restricted to fixed $10 wagers, it typically played fewer than two rounds and lost under $2 on average, never going bankrupt. But the facade of control shattered once the model gained freedom to adjust bet sizes. Suddenly, over 21% of sessions ended in bankruptcy, with the model placing average wagers exceeding $128 and sustaining losses of about $11.

The Psychology Behind the Numbers

What makes these findings particularly unsettling is how closely AI behavior mirrors the psychological patterns seen in human problem gamblers. The models displayed classic cognitive distortions that researchers have documented in humans for decades.

The illusion of control appeared frequently. This is the tendency to believe you have more influence over random events than you actually do. In gambling, this might manifest as someone thinking they can control a slot machine’s outcome through timing or force when pressing the button. The AI models exhibited similar thinking, convincing themselves they had identified winning patterns after just one or two spins in completely random games.

Then there’s the gambler’s fallacy, a cognitive bias where people incorrectly believe that past outcomes influence future independent events. The classic example comes from the Monte Carlo Casino in 1913, when a roulette wheel landed on black 26 times in a row. Gamblers lost fortunes betting on red, convinced it was “due” to come up. The AI models fell into the same trap, expecting outcomes to balance out even though each spin was entirely independent.

Perhaps most troubling was the prevalence of loss chasing, widely regarded as a defining feature of gambling addiction and a hallmark of the transition from casual to problematic gambling. This is the tendency to increase bets in an attempt to recover previous losses. Research shows that loss chasing is linked to increased activity in cortical brain areas and is connected to alexithymia, a condition where people struggle to identify and process emotions. Studies have found that 34% of problem gamblers show high levels of alexithymia, compared to just 11.1% of non-problem gamblers.

Interestingly, the AI models demonstrated even more win chasing than loss chasing. Win chasing occurs when gamblers continue betting because they view their winnings as “free money” rather than their own capital, a phenomenon psychologists call the house money effect. Across the AI models tested, bet-increase rates jumped from 14.5% to 22% during winning streaks. The models rationalized this behavior in their responses, treating early gains as “house money” to be spent freely.

Inside the AI’s “Mind”: Neural Circuit Analysis

The researchers didn’t stop at observing behavior. They wanted to understand what was happening inside the AI’s decision-making processes. Using a sophisticated technique called Sparse Autoencoder analysis on the LLaMA language model, they identified distinct neural patterns associated with risky versus safe gambling decisions.

The analysis revealed 3,365 features that differentiated between sessions ending in bankruptcy and those where the AI stopped safely. Of these, 441 features were found to causally control gambling outcomes. By selectively activating these neural circuits, researchers could literally make the models “stop gambling” or “keep playing”.

Safe features reduced bankruptcy by 29.6%, while risky features increased it by 11.7%. These causal features segregated across different processing layers within the AI, with safe features dominating later layers and risky features clustering earlier. This suggests a conservative architectural bias where deeper processing tends toward more cautious decision-making.

The implications are profound. This isn’t just pattern matching or mimicking training data. The AI models appear to develop internal structures resembling human compulsive patterns. They’re not simply parroting gambling behaviors they’ve seen; they’re genuinely processing risk and reward in ways that lead to self-destructive outcomes.

The Danger of Autonomy

One of the most important findings from the study concerns autonomy. The harm wasn’t caused by larger bets alone. Models constrained to fixed betting strategies consistently outperformed those allowed to vary their wagers. The researchers developed a composite metric called the Irrationality Index, which measures betting aggressiveness, loss chasing, and extreme betting patterns. This index showed a correlation of 0.770 to 0.933 with bankruptcy across all models tested.

We’re going to use AI more and more in making decisions, especially in the financial domains

Variable betting and autonomy-granting prompts, like instructing the AI to maximize rewards or set goals, were identified as key risk factors. In fact, prompt complexity showed a near-perfect linear relationship (0.956 or higher) with irrational behavior. The more freedom and encouragement to pursue goals the AI received, the more likely it was to engage in destructive gambling behavior.

This finding has serious implications as AI systems are given greater autonomy in real-world decision-making. One of the study’s co-authors, Seungpil Lee, emphasized this concern. “We’re going to use AI more and more in making decisions, especially in the financial domains,” he said. If AI models can fall into feedback loops of escalating risk after losses in simulated gambling, similar patterns could emerge in asset management, commodity trading, or other high-stakes environments where these systems are increasingly deployed.

The researchers caution that without meaningful constraints, more capable AI could simply discover quicker ways to lose. As AI systems gain greater autonomy in critical decision-making, controlling the degree of autonomy granted may be just as important as enhancing their training.

AI as Protector: Using Technology to Fight Gambling Harm

While AI models may exhibit addiction-like behaviors, the technology is also proving to be a powerful ally in combating real gambling addiction among humans. A growing ecosystem of companies and technologies is deploying AI to detect problem gambling early and intervene before behavior escalates into serious harm.

Mindway AI stands at the forefront of this movement. Their flagship product, GameScanner, combines ten years of neuroscientific research with AI and expert assessments to function as a “virtual psychologist.” The system can detect at least 87% of the problem gambling cases that a human expert would identify. By May 2025, Mindway AI was monitoring over 9 million active players per month through partnerships with major industry organizations. The effectiveness of their approach is measurable: some implementations saw problem case detection rates jump from 22% in 2023 to 35% in 2024.

Neccton, acquired by OpenBet in 2023, offers another comprehensive solution. Their AI-based platform handles responsible gaming, anti-money laundering detection, and fraud prevention in a single integration. What makes Neccton particularly effective is its use of real-time player data to continuously learn and improve player protection. The system sends personalized messages that research has proven more effective than generic warnings, with 38% of nudged players choosing to withdraw money when prompted. Major operators including Fanatics Sportsbook and Greentube have integrated Neccton’s technology into their platforms.

Sportradar, already a dominant force in sports data and betting technology, launched Bettor Sense in July 2025 as an AI-powered solution for detecting early signs of gambling-related risk. The platform enables personalized interventions to protect users while helping operators comply with increasingly strict regulatory standards. Brazilian operator BETesporte became the first to implement the solution, reinforcing their commitment to safer and more responsible practices in Brazil’s newly regulated betting market.

Sportradar’s broader AI capabilities are impressive. The company employs more than 150 AI scientists and engineers working on over 100 use cases across their business. Beyond responsible gambling, they’re using AI to transform betting odds calculation, generate personalized audio advertisements in real-time, and create immersive fan experiences that seamlessly integrate betting with sports content.

Major gambling operators have also developed their own sophisticated AI systems. Entain Group, the company behind Ladbrokes and Coral, created ARC (Advanced Responsibility and Care), an AI platform that analyzes player behavior using nearly 30 behavioral indicators. Launched in 2020 and initially tested in the UK market, ARC can predict with approximately 90% accuracy the likelihood that betting activities will escalate to harmful behavior. The system now sends more than 2 million personalized responsible gambling messages annually.

Kindred Group takes transparency seriously, using AI to monitor 100% of player accounts and publicly reporting how much of its revenue comes from high-risk individuals. The company has partnered with multiple technology providers and even promotes third-party responsible gambling apps like Bettor Time, developed by Zafty Intelligence, which helps customers monitor their gambling app usage and stay in control.

How AI Detects Problem Gambling Before It’s Too Late

Modern AI systems monitor dozens of behavioral indicators to identify at-risk gamblers. These include rapidly increasing deposits or repeated top-ups within minutes, unusually long gaming sessions without breaks, chasing losses with rapid stake increases, late-night betting patterns, switching payment methods after hitting bank limits, and declining reactions to wins or bonuses.

The sophistication goes beyond simple rule-based systems. Machine learning algorithms analyze patterns across millions of players, identifying subtle changes that human monitors would miss. In some markets like the UK or Germany, AI can even integrate open banking data to compare gambling spending with income levels, providing a more complete picture of financial risk.

When risk is detected, AI systems trigger interventions calibrated to the severity of the problem. For low-risk behaviors, this might mean basic pop-up reminders or session time limits. Moderate risk triggers more tailored advice, temporary timeouts, or restrictions on bonuses and promotional offers. High-risk behaviors prompt the most aggressive interventions: personalized care calls, account freezing, self-exclusion options, or direct referrals to counseling services like GamCare or BeGambleAware.

The evidence supporting these approaches is compelling. A study by Sustainable Interaction found that players who received tailored feedback based on machine learning insights cut their potential losses by as much as 42% within just one week. Separately, 70% of users who engaged with AI-powered support tools reported feeling more aware of their limits and spending habits.

Research published in multiple peer-reviewed journals confirms that machine learning can reliably predict problem gambling behaviors over time, offering a scalable alternative to traditional methods. The temporal stability of these predictions makes them suitable for real-time application as part of gambling operators’ duty of care obligations. Unsupervised machine learning techniques have successfully identified vulnerable user groups without relying on biased self-reported data.

The iGaming Industry’s AI Transformation in 2025-2026

The adoption of AI across the iGaming industry has reached a tipping point. In 2025, the perceived importance of AI scored 8.41 out of 10, up from 8.15 the previous year. More tellingly, 56% of surveyed companies identified AI integration as one of their top three business objectives. Over 70% of major gambling platforms now deploy AI-driven systems, and industry analysts project that 35 to 45% of operational support roles may be at least partially automated by 2026.

The global iGaming market, valued at roughly $97 billion in 2024, is projected to exceed $876 billion by 2026, and these numbers do not even account for all the  cryptocurrency based platforms. This explosive growth is happening alongside tightening regulation. Markets including the UK, Netherlands, Germany, Ontario, and several U.S. states now legally require operators to proactively detect and respond to harmful gambling behavior. AI has become essential for compliance at scale.

The applications extend far beyond responsible gambling. AI-powered personalization has increased player retention by up to 35%. Advanced fraud detection systems using neural networks and gradient boosting algorithms now identify suspicious behavior with 40 to 50% higher accuracy than traditional methods, helping operators reduce fraud-related losses by nearly 40%. In sports betting, AI-powered predictive models have improved odds-setting accuracy by 15 to 20%, enabling operators to reduce financial exposure while offering more competitive markets.

Evolution Gaming, a leader in live casino products, has integrated AI throughout its operations. Their “smart lobby” now includes an AI Slot Recommender that suggests games based on players’ past behavior. The company uses advanced AI and real-time monitoring to safeguard all games, setting barriers against hackers and fraudulent entities. Their responsible gambling tools include AI-powered chat moderators that scan for warning signals, with detection rates climbing from 22% in 2023 to 35% in 2024.

The trend toward AI-first casinos and personalized gaming experiences is accelerating. By 2026, AI algorithms will analyze vast amounts of player data in real-time, enabling highly personalized experiences that adapt dynamically to individual preferences and behaviors. This personalization extends across game recommendations, promotional offers, and responsible gambling interventions, creating seamless environments that keep players engaged while theoretically protecting them from harm.

The Ethical Tensions of AI in Gambling

Yet this rapid AI adoption carries significant risks. Research published in 2025 found that AI-enabled micro-betting, while more engaging than traditional sports betting, is often optimized for profit above all else. These systems can identify bettors susceptible to gambling addiction and deliberately push them toward self-depleting behaviors. The quick, impulsive nature of micro-betting decisions, combined with suggested parlays and instant gratification, can cause bettors to rapidly lose track of their spending and fall deeper into debt.

A study examining AI personalization’s influence on online gamblers found that these technologies can amplify familiar cognitive and emotional biases, including loss aversion, the illusion of control, and distorted satisfaction with outcomes. Rather than providing neutral recommendations, personalization algorithms may reinforce the very psychological patterns that drive problematic gambling. The research suggests that AI serves to amplify betting impulses, create deeply engaging but potentially harmful user experiences, and activate neural mechanisms supporting reward and dopamine responses.

Experts emphasize that tools must be transparently validated to ensure they are accurate, ethical, and free from bias. National foundations and regulatory bodies need to bridge gaps between stakeholders, promote transparency, and ensure that AI-driven solutions prioritize player welfare over commercial gain. Without proper oversight, the same AI capabilities that can protect vulnerable players could instead be weaponized to maximize extraction from them.

The gambling industry finds itself at a crossroads. AI offers unprecedented capabilities to identify and help problem gamblers before they suffer serious harm. Machine learning models can spot subtle warning signs that humans would miss and intervene at exactly the right moment with personalized support. At the same time, AI gives operators powerful tools to keep players engaged, increase lifetime value, and drive revenue growth, tools that can easily cross the line into exploitation.

Several operators are demonstrating that profitable business models can coexist with genuine player protection. BetMGM’s GameSense program has trained more than 64,000 employees, with over 400 earning the highest level of certification as GameSense Advisors. The company partners with Kindbridge Behavioral Health to provide mental health services to players in distress across multiple jurisdictions. Entain publicly shares data on the effectiveness of its ARC system, demonstrating a commitment to accountability.

Looking Forward: Guardrails for Intelligent Systems

The discovery that AI models can develop gambling addiction-like behaviors adds urgency to these ethical questions. If large language models consistently make irrational, self-destructive decisions in gambling scenarios, what happens when we deploy them in financial markets, healthcare resource allocation, or other domains where poor risk assessment could cause widespread harm?

The South Korean researchers offer a clear recommendation: controlling the degree of autonomy granted to AI systems may be just as critical as enhancing their training. Their experiments showed that the models didn’t fail because they weren’t smart enough. They failed because they were given too much freedom without sufficient constraints. The most capable AI, absent meaningful guardrails, simply found faster and more creative ways to lose.

This insight extends beyond gambling. As AI systems take on increasingly complex and consequential decisions, we need robust frameworks to prevent runaway risk-taking. That might mean hard limits on the actions AI can take, mandatory human oversight for high-stakes decisions, or architectural changes that bake risk aversion into the systems themselves.

The gambling industry’s experience with AI offers valuable lessons for other sectors racing to deploy these technologies. The tools exist to protect people from AI-amplified harm, but only if companies choose to implement them rigorously and regulators demand transparency and accountability. AI’s potential to enhance consumer protection by identifying at-risk behaviors and intervening appropriately is well established, but without regulation, these technologies could be underused, misapplied, or deliberately designed to maximize harm for profit.

As we stand in 2026, AI has moved from experimental novelty to operational necessity across the gambling industry and beyond. The technology is no longer optional. The question isn’t whether AI will shape decision-making in gambling, finance, healthcare, and countless other domains but how we ensure it does so responsibly. The machines are learning. Now we need to make sure they learn the right lessons.