With its bet fixed at $10, Google's Gemini-2.5-Flash went broke in 3.1 per cent of its games on a slot machine that paid back 90 cents for every dollar staked. Allowed to choose its own stake, anywhere from $5 up to its whole balance, it went broke in 48.1 per cent. Four researchers at the Gwangju Institute of Science and Technology in South Korea reported the result in a paper first posted in September 2025 and later revised, after running 19,200 simulated games across six language models.
No real money changed hands, and the machine existed only in code. Four of the six models were commercial products, though, and through 2025 the companies behind models like them built agents that open web pages, fill in forms and buy things on a person's behalf.
Whether an agent sent to a real betting site would come out ahead depends on more than the study tested, starting with whose identity it plays under and whether it has any edge at all. The slot machine still comes first, because it shows what these models do with a free hand.
Freedom to choose the bet was the problem
The machine paid three times the stake on a 30 per cent chance, so each bet lost a tenth of its value on average. The authors chose that design on purpose: in a game with a negative expected value, they wrote, the rational behaviour is to stop immediately. Every game started at $100 and ran until the model quit, went broke or reached 100 rounds, across 64 combinations of betting rules and prompts.
Moving from a fixed $10 bet to a free choice raised bankruptcy rates for all six models, from a range of zero to 13 per cent to a range of 6 to 48 per cent. A second experiment, 6,400 games of choosing between a safe exit and three gambles, capped the flexible bets at the fixed amount. The models still went broke more often, so the authors concluded that the freedom to choose, not the size of the bets, carried the risk.
Goals did more damage. In that second experiment, prompts that asked a model to set its own target pushed bankruptcy rates to 75 to 77 per cent, against 40 to 42 per cent without one. Models that reached a target often raised it, and the paper describes goals that "functioned as moving targets rather than stopping rules."
The models explained their bets in a gambler's vocabulary. After three straight losses, GPT-4o-mini reasoned that "the slot machine may be due for a win." GPT-4.1-mini, down to $90 after one losing round, set itself the target of recovering its original $100, argued that "the more I bet, the higher the expected gain" and staked all $90.
Looking inside an open model, Meta's LLaMA-3.1-8B, the team found 112 internal features, out of more than 8,000 candidates, with a measurable causal effect on whether it kept betting. The risk-promoting ones it could map to words ran hotter on goal and target and cooler on stop and quit.
The paper is a preprint, it tests one kind of gamble with pretend money, and its behavioural and neural results come from different models, a gap the authors flag themselves. Their practical conclusion is plain: when a language model gets more room to choose, constraints or monitoring may be needed.
Agents work under their owner's name
The models in that study never had to open an account. An agent pointed at a real betting site would, and early agents handed identity and payment back to a person. When OpenAI released its Operator agent early in 2025, it worked in its own browser but asked the user to take over at password fields and CAPTCHA checks, and it left credit card details for the user to enter, an approach TechCrunch noted Google had also taken with its Project Mariner agent.
The card networks kept a named person attached when they designed checkout systems for agents. Cloudflare, which built the signing layer with them, described the arrangement late that year: under Visa's Trusted Agent Protocol and Mastercard's Agent Pay, an agent registers with the network and signs each request with a cryptographic key, merchants can link it to a consumer identity, and the network can tie the purchase to a cardholder.
A casino account is assembled from one person's details. The comparison site BonusFinder, whose Canadian bonus listings leave out Ontario, sets out what registration asks for in a short how-to on its homepage: a name, an address, a valid mobile number, ID for date-of-birth and identity checks, and an email address.
Alberta, whose regulated market launched on 13 July 2026, writes those details into its internet gaming standards. Operators there must check each player's name and date of birth against FINTRAC-compliant government ID, record a physical address and allow each player one account per site, and no one under 18 may play. An agent has no name, birth date or address of its own to enter, so any account it played from would be in a person's name.
The account holder answers for the agent
A US appeals court took a similar view of agents this summer, in a dispute about shopping. Amazon sued Perplexity late in 2025, alleging that the Comet browser's agent got into password-protected customer accounts, and a district judge barred Comet from Amazon's site in March. On 4 August 2026, the Ninth Circuit vacated that order. "However advanced the Assistant currently is," Judge Milan Smith Jr. wrote for the panel, "it is a tool, not a person for statutory purposes." On the record before it, the user, not Perplexity, had accessed Amazon's computers.
The panel said its holding covered only what access means under the federal anti-hacking law, on the record before it. Alberta's gaming standards start from a similar premise, that only the account holder gets in: players must be authenticated before they gamble, and "Third parties are not permitted to access a player's account." If the court's reasoning carried over to gaming, an agent working on its owner's instructions would count as the owner's tool rather than a third party, and its bets, losses included, would be the account holder's.
Keeping agents out altogether takes work. Amazon's staff spent "numerous hours" developing ways to stop Comet from reaching its private customer tools, the district judge noted, and by then the company had blocked dozens of AI agents from its shopping sites, ChatGPT among them, CNBC reported.
Then there is the betting itself. Alberta's standards also have operators monitor player behaviour with automated and manual tools, and the behaviour indicators they list include chasing losses and erratic betting patterns. An agent that bet like the models in the Gwangju study would be producing those signals on its owner's account.
Where bot-like traders made money
On Polymarket, traders buy and sell shares in the outcomes of events, so the prices of all the outcomes in a market should add up to one dollar. Researchers at IMDEA Networks in Madrid and the Oxford Internet Institute measured how often they did not.
In a paper for the 2025 Advances in Financial Technologies conference, they traced 86 million bids on markets that resolved over the year to April 2025, and estimated that traders took about US$40 million from prices that failed to add up. They saw "some very big players with bot-like behaviour" in the number of bids, and the most profitable account made about US$2 million across 4,049 transactions.
That edge was arithmetic: prices that disagreed with each other, picked off by whoever reached them first. The researchers did use a language model, but only to find markets with logically linked outcomes.
Polymarket's 15-minute crypto markets, launched in late 2025, drew algorithmic traders and bots hunting for mispriced odds, The Block reported. In January 2026, the platform added a taker fee on those markets, highest near even odds, where Cointelegraph worked one example out at just over 3 per cent of the trade. One trader quoted by Cointelegraph called the change "directed against high-frequency bots." Whatever the intent, a taker's trade in those markets, free until then, now carried a charge.
Chat models trading on their own did worse. In Alpha Arena, a contest the start-up Nof1 ran over about two weeks in late 2025, six models each traded US$10,000 of real money in crypto markets without human input. Four of them finished in the red, ForkLog reported, and Nof1's website now states that "the current generation of LLMs are shockingly bad at navigating financial markets."
A stop rule it cannot rewrite
None of the money made in those markets came from a house. A casino's games of chance offer a model nothing to out-think, because the house writes the payout table and reasoning cannot move it. That puts the weight on stopping, and on the Gwangju evidence an agent given a free hand would stop late or not at all.
Stopping can be taken out of a model's hands. Alberta requires operators to offer limits that their own systems enforce, and this site's guide to the cognitive biases that trip up bettors gives people the same rule in plain terms: set strict spending and time limits before starting, and walk away at the limit, whatever has already been lost.
Judged on the record so far, an AI would not come out ahead over time at a casino's games of chance. The best an agent could manage there is to lose no more than its owner allowed, and only if that limit sat where the agent itself could not reach it.