It turns out some large language models behave as if they have a stable risk attitude — a consistent tendency to be either cautious or bold when facing uncertain choices. This mirrors something well-studied in humans: some people reliably prefer a sure thing, while others consistently gamble on a bigger payoff.

Researchers have found that certain LLMs show this same consistency across different scenarios, meaning their risk preferences aren't random noise — they're a repeatable pattern. This likely emerges from the training data and the human feedback used to shape the model, essentially encoding average human risk preferences into the model's behavior.

This has real implications for using AI in high-stakes decisions like finance, medicine, or policy. If a model has a hidden risk attitude, it may subtly push recommendations in one direction without users realizing the model has a consistent lean.