Large language models (LLMs) are the technology underlying chatbots such as ChatGPT. They are often credited with human traits: intelligence, intentions, bias and even beliefs. But do they have their own attitude to risk? That is exactly what Erik Jing Du, a professor of engineering at the University of Florida, set out to find out.
His team will present the paper "Benchmarking Risk Attitudes of Large Language Models" in December at the annual NeurIPS 2026 conference in Sydney. The co-authors are Bowen Sun, Rui Min, Xianyao Li, Yuxi Wang, Ryan Qi Wang and Gilbert Yang Ye.
The researchers started from the premise that LLMs are increasingly used for real-world decision-making, sometimes in military contexts. They were interested in how models behave in situations of risk and uncertainty. "For several years, research has chased high model performance," Du says. "We look at a different dimension: risk appetite, or the model's attitude."
To test whether models have stable traits independent of training data, the team forced them to "forget." The models were given the same complex tasks, for example medical triage (assigning patients by urgency of care). Then the patient information was erased again and again and the questions were repeated. The model could not rely on the previous dialogue, and each time the situation was new.
After hundreds of repetitions, the models' assessments remained surprisingly consistent. According to Du, this suggests that LLMs are not merely probabilistic machines but systems with some internal mechanism that ensures consistency.
The researchers then tested how modern models behave in three complex tasks: piloting a drone in a windy, crowded city, choosing an investment portfolio of stocks and bonds in a financial environment, and medical triage. It turned out that a model's risk appetite persists from task to task: if a model took risks in one, it took risks in another.
But there was another difference. "We saw significant differences before and after training models with human feedback," Du says. "That means large companies use training processes that finely tune the model's personality. Since each does it its own way, we see different personalities."
Human experts also took part in the work. They were shown logs of the models' decisions - records of their reasoning - and asked to assess the level of risk without knowing the model's name. "It was like a Pepsi Challenge," Du explains. "We asked: is it risky or not? And we found a significant gap between human perception and the model's actual risk appetite."
Du hopes that a risk check will become a standard step before models are deployed. Currently, he says, this is not done even for advanced systems, which also surprised the paper's reviewers. "You may want to use a cautious model, for example for air traffic control or in medicine, but in practice deploy a risky one. That is dangerous," he warns.
For now this is research rather than a finished product. The work is intended to help understand how AI makes decisions when the outcome is uncertain and something has to be sacrificed.