Recent studies, including research by Sun et al. (2026), have analyzed the risk-taking behavior of large language models across various scenarios, such as drone navigation, hospital triage, and investment portfolio management. Models were evaluated on their ability to assess danger and make decisions, with most exhibiting consistent relative risk profiles.
The observed hierarchy of risk-taking behavior among popular models, from least to most risky, is Claude Sonnet 4.5, DeepSeek V3.2, Gemini 3 Pro, Qwen3 Max, and GPT-5.2. Grok 4 showed inconsistent behavior, acting as the most risky in drone tasks but the least risky in others, while also demonstrating high variance in danger assessment.
Research indicates that while prompts and language settings can influence risk-taking, the core risk attitudes are largely shaped by pre-training and fine-tuning processes focused on safety and utility. Verbal assessments of risk were found to be largely independent of actual behavioral choices in models like Gemma-2-9B-Instruct and Llama-3.1-8B.
The study of risk in LLMs involves distinct forms of internal processing, including danger assessment, behavioral choice, and verbal description. These forms are often linked to specific tasks and attention mechanisms, suggesting that risk-taking is a byproduct of model alignment rather than a primary design goal.
