Moral Framing and Decision-Making Under Risk in Large Language Models

Abstract

As AI gradually manages high-stakes moral decisions, understanding its consistency is crucial. The study examined framing effects in five LLMs: Gpt-4.1-mini, Gpt-5-nano, Claude-3-Haiku, Gemini-2.5-flash, and Gemini-2.0-flash, testing 1,250 decisions across three morally challenging scenarios. The research investigated if linguistic framing alters risk preferences and if mathematical reasoning reduced bias. Results revealed that AI models choose the moderate option, which weakens framing effects compared to not morally challenging scenarios (Payne 2025). The most primary findings was that even though logical reasoning was processed and mentioned in the rationale, AI models still exhibited significant framing effects. This suggests that AI reasoning is influenced by biased choices rather than making the final decision logically. Since models are trained differently, and showed significant inter-model variability they behave inconsistently. Therefore, highlighting the necessity for training models to be unbiased during alignment level in systems that are automated.