Humans are highly susceptible to framing manipulations in intertemporal decision making—choices that involve tradeoffs between immediate and delayed rewards. Classic biases such as the magnitude effect, in which larger reward amounts increase patience, have been attributed to self-control, reward system activation, and other cognitive mechanisms. Here, we show that large language models (LLMs) exhibit similar sensitivities to decision framing, including the magnitude, sign, and hidden-zero effects. Unlike humans, LLMs discount delayed rewards exponentially rather than hyperbolically and do not exhibit a decimal effect, suggesting that their behavior is not merely inherited from descriptions of decision-making phenomena in their training data.
Analysis of LLM embedding spaces, which encode semantic knowledge, revealed that large monetary amounts are represented more closely to words related to delay and the future. This suggests that framing biases in LLMs arise from semantic proximity of choice options to temporal concepts. Together, these findings introduce a conceptual framework in which linguistic structure shapes decision-making, raising the possibility that human framing effects partly emerge from the organization of choices in semantic space.
Ian Ballard is an Assistant Professor at the University of California, Riverside. He is affiliated with the Cognitive and Cognitive Neuroscience area of the Psychology Department and the UCR Neuroscience Graduate Program. He uses an integrative approach in which insights from statistical learning, neuroscience, and cognitive psychology inspire novel experimental probes of the relationship between goals and learning. His research has been funded by the National Science Foundation and National Institutes of Health.