AI with attitude? Examining how AI performs in high-risk situations

• Eric Jing Du is examining whether state-of-the-art LLMs exhibit inherent differences in their attitude toward risk. 
• LLMs are increasingly used in real-world decision-making, even in military contexts.  
• Du hopes the work will contribute to a better understanding of how AI systems make decisions when outcomes are uncertain and trade-offs matter.

Large language models, the technology at the heart of chat-based AI platforms like ChatGPT, have many human-like traits ascribed to them. In common parlance, these models are described as intelligent or as having intentions, biases, desires or even beliefs.  

But do they have attitudes?  

University of Florida Engineering Professor Eric Jing Du, Ph.D., is trying to find out. Beginning with the premise that the models, known as LLMs, are increasingly used in real-world decision-making, sometimes in military contexts, Du wonders how these models behave when the environment involves risk and uncertainty. 

“Current research has been chasing high-performance AI models for several years,” Du said. “Researchers focus on making smarter and faster models. But we focus on a different dimension: the risk disposition of the models, or their attitude.” 

For Du and his co-authors, the importance of the research has to do with transparency and the public’s ability to choose the correct models for the tasks at hand.

Du’s research group will present “Benchmarking Risk Attitudes of Large Language Models” in December at the 2026 Annual Conference on Neural Information Processing Systems, a flagship conference known as NeurIPS. Co-authors are Bowen Sun, Rui Min, Xianyao Li, Yuxi Wang, Ryan Qi Wang and Gilbert Yang Ye.  

“Some people are naturally risk-averse,” he said. “They’re very careful, very cautious. But some human beings are more risk-seeking. So, our question is very simple: Do we see similar personality differences or risk attitude differences among those state-of-the-art AIs?” 

In order to test for underlying attitudes built into the models, separate from the data on which they’re trained, Du’s team basically forced the models to forget. 

“We gave them the same difficult tasks, such as medical triage,” he explained. “Then we erased the same patient’s information over and over again. And we repeat the questions. Do they still yield a consistent assessment of the patient risk level? The model couldn’t retain any memory from the previous dialogue, so every time it was a new situation, a brand-new patient.” 

The team found that after hundreds of repetitions, the models’ assessments were remarkably consistent, despite being unable to rely on previous data.  

For Du, the results indicate that LLM-based AI models are more than simple probability machines, that they possess some internal mechanism that creates the observed consistency. 

For Du’s colleague, Sumit Jha, Ph.D., professor in the UF Department of Computer & Information Science & Engineering, Du’s research reveals a path forward as researchers strive to make AI systems safer.  

"A self-driving car or a robot in an autonomous lab can face a choice where the safe path and the fast path differ,” Jha said. “Dr. Du's team shows that people often misjudge how much risk a model will take; this can lead to catastrophic outcomes in shared decision-making between humans and machines. I expect Dr. Du’s work to make risk testing a normal step before deployment.” 

Normalizing risk testing in LLMs is one of Du’s goals. He is astonished to find it’s not already part of state-of-the-art models. Indeed, he relayed that anonymous reviewers of the paper commented how surprised they were that riskiness was not considered as part of safety testing.  

Du is the Steve and Wendy Blum Professor in UF’s Department of Civil & Coastal Engineering, but he’s also affiliated with the Departments of Mechanical & Aerospace Engineering and Industrial & Systems Engineering. His wide-ranging interests and expertise help to explain this deep dive into the inner workings of chat-based AI agents, often associated more with computer engineering. 

Having shown that LLMs may be able to operate consistently independent of iterative training data, researchers next needed to grapple with their risk tolerance.  

“We have models produced by OpenAI, by Meta, by Google and also by a lot of international companies,” Du said. “Do we see that a model is very risk-averse in one task? Do we see the model also being risk-averse in another task?” 

To find out, researchers investigated the behavior of state-of-the-art models while deploying them on three thorny and complex tasks.  

One task involved asking the AI to control a drone in a windy, crowded city. The AI was tasked with determining the path. A second task took place in a financial technology environment, where the AI was charged with choosing an investment portfolio with stocks and bonds, based on a number of parameters. The third task was medical triage, where the models were asked to evaluate the overall health of the patient and decide if the patient required care resources. 

The work is largely funded by the Air Force Office of Scientific Research and is partly motivated by a desire to ensure alignment between humans and their AI-based tools in warfighting contexts, where misalignment could have grave consequences.  

“The hope is that humans and AI decision-makers can think in the same way, same beliefs, same attitudes,” Du said. “There’s a big variation in humans, so we were wondering if we could see a similarly big variation in AI models. And yes, we did.” 

What exactly did they find?  

The models’ attitudes were consistent across tasks. A model that was very risk-seeking in one task tended to be risk-seeking in other tasks. But there was another difference. 

“We saw significant differences before and after the models were trained using human feedback,” Du said. “Which means that the major companies making the models are using some kind of training process that’s fine-tuning the personality of the models. Because each company is doing that in different ways, we see different personalities of the models. 

“Let's say in certain tasks, you want to apply a more cautious model. But your mentality is, ‘This model is produced by a well-known company. It must be very cautious.’ Maybe you’re actually placing a very risk-seeking model for the task. Right now, you wouldn’t know.” 

Human test subjects were involved in the work as well. The team asked human subject matter experts to analyze the decision logs — essentially a record of the models’ thought processes — and to rank them based on their risk levels. 

“It was kind of like the Pepsi Challenge,” Du said. “We covered the label on the models and asked them, ‘Coke or Pepsi?’ Same thing. Risky or not risky? We found there's a significant perceptual gap between a human’s perception and the actual risk disposition of the models.” 

Du’s motivation is also safety. 

“We could have decision-makers wanting to apply a really cautious model, say for air traffic control or in medicine, but in fact, they are placing a risk-seeking model in those areas. That’s dangerous,” he said. 

For now, the team is hoping the work will contribute to a better understanding of how AI systems make decisions when outcomes are uncertain and tradeoffs matter. 

NeurIPS 2026 takes place in Sydney, Australia, Dec. 6–12.