Money illusion in large language models. An exploratory replication study

As Large Language Models become increasingly integrated in daily life, understanding the extent to which they may reflect behavioural tendencies embedded in the training data becomes a relevant issue. In this paper, we conduct an exploratory replication study on the occurrence of “money illusion” across six Large Language Models: GPT-5, GPT-5.1, DeepSeek-V3.2, Grok 4.1, Sonnet 4.5, and Gemini-3-pro-preview. Specifically, we replicate the survey questions of Shafir, Diamond, and Tversky (1997) and report on the models’ responses across twelve scenarios covering earnings, transactions, contracts, and mental accounting. Our exploratory evidence tends to suggest that Large Language Models generally base their responses on real terms rather than nominal ones, particularly in contexts that require objective evaluations, even though there are some exceptions. However, it also tends to suggest that their responses more frequently aligned with those of the human participants in the original study when more subjective evaluations were involved. While a more systematic investigation based on a larger dataset is required to reach more definitive conclusion, we nonetheless hope that our results could be useful for future research analyses.