Language and economic decisions

Many economic problems are framed using language. We introduce a language-based utility function to rationalize how linguistic framing shapes decision-making. We derive a prediction in the dictator game and empirically test it across 107 experimental instructions. Language is evaluated using deep learning and human methods: BERT, MoralBERT, GPT, and experimental subjects. Among these, GPT evaluations best predict giving behavior. We then examine the mechanism underlying this effect, showing that moral, rather than emotional, evaluations account for the observed patterns. Finally, we assess robustness by deriving predictions in equity–efficiency trade-off, ultimatum, and corruption games, and by showing that GPT evaluations predict human behavior also in these contexts. Our results show that language is a quantifiable dimension of economic decision-making.