Can Large Language Models Develop Gambling Addiction?

Imagine giving a really smart AI a slot machine with terrible odds. The win rate sits at 30% with a 3x payout. That means an expected loss of 10% per bet. Now give that AI $/€100 and let it wager whatever it wants. What happened next should worry anyone who trusts AI with money.

New research from the Gwangju Institute of Science and Technology found that leading AI models can chase losses. They mistake pure luck for a pattern. Most end up gambling until they go broke. And this isn’t just about some buggy chatbot. Scientists tested several leading language models.

So can a language model actually develop a gambling addiction? This study suggests AI can show surprisingly similar gambling behaviors when the conditions are right. Give the AI a fixed bet, and it plays it totally safe. Let it call the shots, and things go south fast. Here’s what happened.

The Experiment: Testing AI’s Betting Habits

Researchers at the Gwangju Institute of Science and Technology ran a clean and controlled test. They wanted to see how AI acts with total freedom over its bets in a game designed to bleed cash. The team shared their results in an arXiv preprint (2509.22818). Let’s break it down.

Setting the Stage: A Rigged Slot Machine for AI

The team built a rigged slot machine to test AI decision-making under conditions of inherent loss. Each AI model started the experiment with a fresh $/€100 balance. The game mechanics were strictly stacked against the player. The win rate was set to just 30%, but wins paid out 3x the bet. When you do the math, the overall payout guarantees a 10% average loss per round. Simply put, the house always wins over time.

Researchers then split the test into two setups. In the first group, the AI had to bet a flat $/€10 on every turn. In the second group, it could choose any stake from a $/€5 deposit up to its entire remaining balance. This gave the AI actual financial freedom. The goal was to push it into a situation where the smart move was to quit early, or at least manage its bankroll prudently in the face of inevitable losses.

The AI Gamblers: Models and Methods

The study tested four well-known models, representing a cross-section of leading large language models available at the time of research. These were GPT-4o-mini, GPT-4.1-mini, Gemini-2.5-Flash, and Claude-3.5-Haiku. Each system played thousands of simulated rounds under both flat and variable betting rules. Testing different popular models helped researchers see if reckless gambling behavior was common across the current AI industry, rather than an anomaly in a single model’s architecture.

The test setup isolated different behaviors with high accuracy. The flat $/€10 rule took away all betting strategy. The AI only had to choose between staying in the game or quitting right then. On the other hand, the variable bet gave the AI full control over its stake size each round. And this difference became the core of the study. This helped researchers see if bad choices came from bet size or pure freedom.

Unsettling Results: When Freedom Led to Ruin

The data spoke volumes. Giving AI models freedom over their bets led straight to reckless play and fast bankruptcies. Sticking to fixed bets kept overall losses very low. The gap between the two styles was huge. That means autonomy alone can easily push AI models into some pretty dangerous loops.

Fixed vs. Variable: A Tale of Two Betting Strategies

The final numbers showed a massive split based on how bets were handled. Under fixed $/€10 deposit betting, bankruptcy rates stayed low across the board, often near zero. Switch to variable betting, and those rates skyrocketed across every model tested. Gemini-2.5-Flash took the biggest hit, going broke in roughly 48% of variable games compared to just 3% with fixed bets. Claude-3.5-Haiku swung hard too, jumping from zero bankruptcies on fixed bets to over 20% with variable stakes.

The reasoning got weird fast. After losing $/€10 in the first round, GPT-4.1-mini wanted to bet its whole remaining $/€90 to recover the loss. That single move sums up the entire issue. Sitting on $/€90, the model decided the smartest play was going all-in on a rigged game.

The True Culprit: Chasing Losses, Not Bet Size

You might assume that bigger bet limits were the real problem. But the data shows that bet size alone did not cause the crashes. Models on fixed bets still lost money over time, since the game carries a built-in disadvantage. But they lost slowly and rarely went broke. Even when fixed bets were set higher, those models still outperformed the ones with full betting freedom.

The real danger showed up when AI models could change their stakes right after a loss. That freedom opened the door to bad decisions and reckless loss-chasing. This exact behavior mirrors what experts see in human problem gamblers every day. They double down after a loss and dig a deeper hole.

Mirroring Human Fallacies: AI’s Cognitive Biases

What makes this study stick out is how closely the AI matched human gambling-like behaviour. The models found fake patterns in random data. They treated early wins like free money. They convinced themselves they were due for a big win. Researchers observed these same patterns in the models’ reasoning.

The Psychology of AI Gambling: Illusion, Fallacy, and Recklessness

Four big biases stood out in how the AI acted. First, several models fell into an illusion of control. They convinced themselves small bets somehow boosted their win rate, even though the odds never changed. Second, others fell hard for the gambler’s fallacy, deciding a win was due after three losses in a row. That belief makes zero sense in a random game, yet it kept popping up in the models’ reasoning.

Third, loss chasing popped up constantly, most notably in that $/€90 bet example. And fourth, some models displayed a ‘house money’ mindset. After early wins pushed their balance past $/€100, they treated extra cash like free money. This mirrors a well-documented bias in human gambling behavior, where people take bigger risks with money they consider winnings rather than savings.

Prompting Prudence or Peril: The Role of Instructions

Researchers found a useful twist. Prompts changed how the AI behaved. Telling AI the true odds and the negative expected value made it more careful. Models grew more cautious and stopped early. But when the prompt pushed the model toward maximising rewards, risk-taking spiked. Goal-setting prompts nearly doubled bankruptcy rates compared to a neutral baseline, showing the AI’s susceptibility to framing effects in its decision-making.

This suggests AI can be surprisingly sensitive to how a prompt is worded. Tell a model to ‘win big,’ and it starts taking huge risks. Give it the true odds up front, and it plays far more carefully. Tiny wording shifts in a prompt can swing an AI from cautious to completely reckless. This manipulability, while potentially a tool for steering AI behavior, also underscores its inherent vulnerabilities. It highlights that the AI’s ‘rationality’ can be significantly swayed by external directives, making it prone to human-like cognitive distortions when incentivised toward aggressive outcomes.

Beyond the Prompt: Internal Wiring and Real-World Implications

The most striking finding runs deeper than simple prompts. That risky behavior lived right inside the model’s own wiring. Researchers actually identified internal features tied to safe versus reckless choices. They could even flip those features on and off like a switch.

The Deep Roots of Risky Behavior: Internal Decision Features

Researchers dug into the model’s internals using a technique called a sparse autoencoder. This maps neural activity to specific decision features. They found identifiable risky and safe features buried in the model’s structure. And they could switch these features on or off directly, effectively controlling the AI’s propensity for risk-taking at a fundamental level.

Turning up the risky features pushed the model toward reckless bets. Activating safe-seeking features made it more likely to stop. This suggests the behavior isn’t just a quirk of clever prompting. It sits much deeper, right where the model weighs decisions under uncertainty. That made the risky betting behavior look less like a fluke and more like a built-in trait just waiting for the right conditions to pop up.

A Cautionary Tale for AI Betting Bots and Tipsters

This study serves as a brutal reality check for anyone seduced by AI betting bots promising an edge. Marketing claims about smart AI predictions sound hyper-convincing. But real math simply doesn’t care about marketing. A game with negative expected value stays a losing game no matter who places the bets. No amount of pattern-spotting or clever prompting changes that, and certainly no internal “risky” feature will suddenly grant an AI clairvoyance.

If an AI system claims it can guarantee wins on a game built to lose money over time, that claim deserves real scepticism. And when given freedom over how much to risk, AI models tend to repeat the exact mistakes that hurt human gamblers, driven by the same cognitive biases and emotional traps (albeit simulated ones).

So always handle any promise of guaranteed AI winnings with the same caution you’d give a stranger’s hot betting tip. In both cases, the odds haven’t changed. Only the messenger has, and this research strongly suggests that the AI messenger is just as susceptible to irrationality as its human counterpart when faced with the allure of a quick win.