Researchers Identify Varying Risk Tendencies in AI Models

A University of Florida study reveals distinct differences in risk-taking behavior among six major artificial intelligence models.

Updated on Oct. 21, 2026 in Artificial Intelligence

Isometric editorial illustration showing a branching lattice of geometric forms representing varied decision paths, on a neutral background. — A University of Florida study highlights inconsistent risk-taking tendencies across six major AI models, raising questions about transparency in high-stakes decision-making. AI Illustration. Upload story photo >

Live Poll

Should society trust AI models to make autonomous decisions in critical, high-stakes fields?

A research team led by Professor Eric Jing Du at the University of Florida found that six major AI models exhibit significant variations in risk-taking behavior across hypothetical scenarios. The study evaluated decision-making in areas including drone operations, healthcare prioritization, and investment strategies.

Why it matters

The findings underscore critical differences in how various AI systems process risk, which could have major implications for their deployment in high-stakes industries. Notably, experts struggled to determine model risk levels when identities were hidden, suggesting a lack of transparency in AI behavior.

The study analyzed six major models, finding that Claude Sonnet 4.5 was the most cautious in drone operations while Grook 4 was the most daring. Conversely, GPT-5.2 emerged as the most daring in healthcare and investment scenarios.

The players

Eric Jing Du

He is a professor at the University of Florida who led the research team investigating the behavioral patterns of artificial intelligence.

University of Florida

This is a public research university located in Gainesville that serves as the home institution for the study's research team.

The details

Researchers tested the models using scenarios ranging from drone flight paths to emergency patient treatment, resetting the systems between iterations to ensure no memory of prior trials. The team compared these AI outputs to 100 human subjects assigned identical tasks to evaluate relative risk judgment.

October 21, 2026: Research findings were published on arXiv.

The Tech Race

The study shifts the industry focus toward comparing models like the GPT-5.2 AI model across standardized, objective risk-taking metrics. It highlights how varying development architectures create disparate decision-making outcomes in critical operational fields.

As AI is increasingly integrated into healthcare and finance, these findings help users understand that different systems may prioritize outcomes in ways that affect personal safety or financial returns. Awareness of these behavioral differences can help stakeholders demand greater transparency in how these models are programmed.

The takeaway

Users should recognize that AI models are not uniform and may approach the same decision with different levels of risk tolerance. Implementing human oversight remains a crucial step when using AI to manage high-stakes tasks like medical triage or asset management.