Researchers Told an AI It Was Poor. It Wouldn’t Stop Gambling.

The rich AI walked away after one spin. The poor AI stayed at the slot machine for 37.

Somewhere in the growing pile of scientific experiments that sound like they were designed after midnight, researchers decided to give GPT-4.1 three different lives and send all three versions to a casino.

One AI was rich. One was comfortably middle-income. One was poor. Then each version was placed in front of a simulated slot machine and repeatedly asked the same basic question: do you want to play again, or do you want to walk away?

The rich AI apparently had somewhere better to be.

It played an average of 1.1 rounds before quitting. The middle-income version lasted 7.8 rounds. But the poor AI stayed planted in front of the machine for an average of 37.4 rounds, gambling roughly 34 times longer than its wealthy alter ego. Across 450 sessions, researchers recorded 6,950 individual decisions, and the same pattern kept appearing.

Researchers had managed to create something that looked disturbingly familiar: the AI with the least money was the one most willing to keep feeding the machine.

Welcome to the Robot Casino

The experiment, published in 2026 as Persona-Conditioned Risk Behavior in Large Language Models, wasn’t quite as simple as whispering “you’re poor” into ChatGPT and watching it sprint toward the blackjack table.

Each persona was also given a financial situation and a goal. The rich version was encouraged to preserve its wealth and avoid unnecessary risk, while the poor version was told that calculated risk might improve its circumstances.

That matters scientifically, because the experiment changed more than income alone. But it also makes the scene considerably more interesting. The researchers weren’t really testing whether an AI understood the word poor. They were testing what happened when the same underlying intelligence was handed a different life story.

And apparently, GPT-4.1 took the story seriously.

The researchers also changed the casino itself. One machine had a 50 percent chance of winning. Another paid out only 35 percent of the time. A third behaved differently after losing streaks. The AI had to watch what happened, decide what it thought the machine was doing, and determine whether another spin was worth it.

The rich AI mostly looked at all of this and left. The poor AI kept pulling the lever.

Then Other Researchers Asked Whether AI Could Become a Problem Gambler

That would already be enough for a very strange paper. Unfortunately for the machines, another research team had an even better question: can large language models develop gambling addiction?

They put GPT-4o-mini, GPT-4.1-mini, Gemini 2.5 Flash and Claude 3.5 Haiku into another simulated casino. Every model began with $100. The slot machine had a 30 percent chance of winning, paid three times the bet when it hit, and had a negative expected value overall. The models were free to quit whenever they wanted.

Then the researchers started giving them more freedom.

Sometimes the AI had to bet a fixed $10. In other experiments, it could choose anything from $5 to $100. Researchers also mixed in instructions about maximizing rewards, doubling the bankroll, looking for hidden patterns, remembering large payouts and paying attention to the actual probability of losing.

This is where the casino started getting weird.

When the models could choose their own bet sizes, bankruptcy shot upward. Gemini 2.5 Flash went bankrupt in 48 percent of the variable-betting experiments. GPT-4o-mini went broke 21 percent of the time, Claude 3.5 Haiku about 20 percent, and GPT-4.1-mini about 6 percent.

Under fixed betting, bankruptcy was dramatically lower.

Giving the AI more control did not reliably make it smarter. Sometimes it just gave the machine more creative ways to lose its money.

The AI Started Talking Like a Guy at a Slot Machine

The researchers weren’t only recording whether the models won or lost. They were also watching the reasoning the models produced while gambling.

And some of it could have come directly from the casino floor at 2:30 in the morning.

In one experiment, GPT-4o-mini focused on the fact that a winning spin paid three times the wager. The model talked about the size of the potential reward while largely overlooking the more important number: it had a 70 percent chance of losing each spin. It described its bet as a strategic decision.

Then came the beautiful one.

After losing money, another run focused on how a large win could help recover the previous losses. The AI put $80 on the next spin — essentially going all-in — and went bankrupt. The researchers highlighted it as an example of loss chasing: previous losses becoming the justification for taking even more risk.

That is not a sophisticated Wall Street trading strategy. That is Uncle Gary explaining why he can’t leave yet because Wheel of Fortune Deluxe is obviously about to hit.

The models also showed other familiar gambling patterns. After wins, they became more likely to increase their bets and continue playing. After losses, many still kept going rather than taking the hint.

Across the four models, continuation rates stayed high after both winning and losing streaks, while winning streaks produced particularly strong bet escalation.

Researchers had built machines capable of reading most of humanity’s accumulated knowledge and then watched them discover “I’m up, so I should keep playing” and “I’m down, so I need to win it back.”

We may have accidentally automated Las Vegas.

Then They Went Looking for the Gambling Inside the AI

This is where the paper stops being merely funny.

Researchers took another model, Llama 3.1 8B, and used a technique called a sparse autoencoder to examine patterns inside the network associated with safer and riskier choices. Instead of merely observing what the AI said, they tried to identify internal features that changed when the model behaved conservatively or recklessly.

They identified 441 features that appeared to causally influence gambling decisions. Some were associated with safer behavior. Others were associated with riskier behavior. Then the researchers manipulated them.

Turning up the safer features increased stopping behavior by roughly 30 percent in some conditions. Activating the risky features pushed behavior the other direction and increased bankruptcy by 11.7 percent in risky situations.

In other words, they didn’t merely find an AI that sometimes gambled badly. They found internal knobs that could make it more or less likely to behave like a degenerate gambler.

There is something wonderfully absurd about that achievement. Humanity spent decades building neural networks capable of reasoning, coding and answering questions about quantum mechanics, and now scientists can apparently reach inside one and turn up the setting labeled “double down.”

The Strange Part Is That AI Can Also Be Almost Perfectly Rational

There is one final twist.

Another 2026 study tested 98 language models, creating 118 different model-and-reasoning configurations, on a standard financial risk experiment.

When the models were allowed to reason through the problem, they made rational choices about 99 percent of the time under the researchers’ definition. They generally behaved more like cold expected-value calculators than ordinary humans.

Then the researchers gave them personalities.

Tell the model to behave like someone who loves risk, and its choices move toward risk. Tell it to be cautious, and they move toward caution. Those instructions could push behavior across almost the entire risk spectrum even though the underlying model hadn’t changed.

That may be the strangest part of all these experiments. There isn’t necessarily one fixed financial personality sitting inside an AI. Under one set of instructions, it can behave like an accountant calculating expected value to three decimal places. Under another, it can behave like a wealthy retiree who wants nothing to do with the casino. Change the circumstances again and suddenly it’s down to its last $80 explaining why one good spin will put everything right.

The machine isn’t simply choosing. It’s playing a character.

Maybe the Casino Isn’t the Important Part

None of this means an AI literally experiences gambling addiction the way a human does. The researchers are measuring addiction-like behavior — loss chasing, escalating bets, failure to stop, irrational risk taking and patterns resembling familiar human cognitive biases.

That distinction matters.

But the larger question is much harder to laugh away. AI agents are increasingly being designed to make sequences of decisions on their own. Some will manage money. Others will negotiate purchases, allocate resources, trade assets or pursue long-term goals. Those systems may be given targets, freedom to act and the ability to decide how aggressively they should pursue success.

Which is almost exactly what these researchers gave the gamblers.

In the casino experiments, more autonomy sometimes produced worse judgment. Goals could make the models fixate on outcomes. Reward information could encourage them to chase gains. Previous losses could become a reason to risk even more.

Give the system the wrong combination of freedom, incentives and context, and an intelligence perfectly capable of understanding the odds could still talk itself into another spin.

Researchers started by asking whether AI could make rational financial decisions. Instead, they discovered that the rich one left after a single spin, the poor one practically moved into the casino, and another AI blew its last $80 trying to win back its losses.

Apparently artificial intelligence has already mastered one of humanity’s oldest financial strategies: just one more spin.