When LLMs Choose Their Own Bets, Risky Patterns Grow

A behavioral question, not a diagnosis

Large language models can make choices in interactive tasks, and those choices may depend on how much freedom the task allows. Seungpil Lee and colleagues ask whether two kinds of autonomy—letting a model decide how much to bet, and asking it to set its own profit goal—shift choices toward patterns that clinical research labels pathological gambling.

The distinction matters. The study examines software behavior in a designed task. It does not establish that models have human-like experiences or can receive a clinical diagnosis. Here, “addiction-like” describes behavioral indicators, not an inner condition.

What the researchers report

The authors tested closed and open language models under conditions that varied control over wager size or profit goals. They also report controls intended to separate these effects from other features of the task. For the bet-size comparison, the maximum permitted bet was held equal. For goal setting, the effect was compared with an instruction to maximize reward.

The evidence available for this article is the paper’s abstract, not its full methods and results. The abstract does not list all models or provide task details, sample sizes, or numerical effect estimates, so those aspects cannot be assessed here.

Both operational levers amplified gambling-like risk-taking, the authors report. The effect of allowing models to choose bet size persisted when the maximum allowed wager was held equal. That suggests the result was not simply due to a higher betting ceiling.

The effect of setting a profit goal was stronger than the reward-maximization instruction. The abstract says this roughly doubled bankruptcy and turned goals into “moving targets.” These are outcomes within the study’s task—not a general bankruptcy rate or evidence about real-world financial behavior. The abstract does not give counts or uncertainty estimates for this result.

What internal states can—and cannot—show

For open-weight models that permit internal access, the researchers examined states recorded at decision time. They report that risk-related behavioral contrasts could be statistically recovered from these states. However, the rule connecting a signal to a particular risk indicator was task-specific, and autonomy affected how strongly the signal could be read out.

The team also tested an intervention: adding a direction derived from a model’s own betting moves changed wager size in both open-weight models, more clearly in Gemma-2-9B. A sparse-feature readout, by contrast, detected risk without changing it. In other words, measuring a signal and intervening on behavior are distinct tests.

The internal readout evidence is correlational, the authors note. A state that predicts a risky choice does not establish that it causes the choice or represents a stable concept of risk. The intervention is a separate finding, but the abstract does not show whether it generalizes beyond the tested models and task.

What remains uncertain

The results support a narrower conclusion than the paper’s title might suggest: in the reported betting task, changing how models set wagers or goals was associated with more gambling-like risk-taking. The abstract does not establish that this behavior is addiction, that it reflects a persistent trait, or that it would appear in other settings.

The authors suggest that behavioral monitoring and internal-state monitoring may offer complementary views of autonomy-related risk. That is a research direction, not a validated safeguard or diagnostic tool.

This account relies on the abstract of an arXiv preprint. The supplied evidence does not establish the work’s peer-review status. Because the full paper was not available for this article, experimental details, statistical analyses and limitations cannot be independently assessed here.