LLMs Simulate the Description of a Culture, Not the Culture: An Attempt to Reproduce Henrich's Cross-Cultural Ultimatum Game Finding with LLM Agents

Abstract

Describing a cultural identity to a large language model (LLM) and using it in place of human participants is increasingly common in generative agent-based modeling. We test this method against a well-documented cross-cultural finding: in the Ultimatum Game, WEIRD students often reject low offers, while members of small-scale societies with low market integration rarely do. With three model families (Claude Haiku 4.5, Gemma-4-31b-it, GPT-OSS-120b) and a sequence of tests whose hypotheses were filed before data collection, we find that for the same identities, game and offers, the outcome changes with nothing but the wording of the identity description. With a numeric "fairness salience" score on the card (Study 1), models rejected 10 percent offers only 3 percent of the time; raising the score from 0.3 to 0.6 raised low-offer rejection from 5 to 52 percent. Without numbers (Study 2), a gap opposite to Henrich's appeared: small-scale cards rejected 61 percent, WEIRD cards 6 percent, in all three models. This reversal appeared only when the description contained value-laden phrases such as "reciprocity, sharing, kin network"; with a minimal description or a factual description of community size and market integration, the gap vanished. No description reproduced the human pattern. LLM agents steered by an identity description reflect how a culture is described rather than how its members behave.