Marketing for a More Sustainable World

Products with a small material diversity (fewer different types of materials) are almost always more eco-friendly because they reduce the use of nonmaterial resources (i.e., they consume less energy during production). Yet, seven pre-registered experiments reveal a paradox: While consumers (almost always rightly) perceive a small material diversity as more eco-friendly when materials are new, this perception reverses when the materials are recycled, consumers (almost always erroneously) perceive a large diversity of recycled materials as more eco-friendly. This reversal occurs because consumers focus on new material resource use (i.e., the use of new materials) instead of the actual driver of eco-friendliness in this context: Nonmaterial resource use (i.e., the use of energy during production). Specifically, greater material diversity increases the perceived total material mass of a product. For new materials, this increased mass is perceived as less eco-friendly because it increases new material use. In contrast, for recycled materials, which are perceived as substitutes for new ones, the same increase is seen as more eco-friendly by implying that more new materials have been avoided. This research advances the extant understanding of how consumers interpret frequently salient quantity information in the marketplace and the underlying drivers of consumers' eco-friendly consumption decisions.

Large language models (LLMs) are increasingly shaping sustainability decisions as both descriptive and prescriptive systems. They may exacerbate ineffective and inequitable climate action through these roles by describing sustainability through synthetic data and prescribing sustainability through personalized recommendations. In Study 1, we examine LLMs in their descriptive role by comparing self-reports of 15 sustainability behaviors from about 2,000 U.S. respondents, stratified across five demographic dimensions (education, income, political ideology, race, and sex), with about 30,000 matched synthetic responses generated by fifteen widely used LLMs. We find systematic distortions in LLM-generated data: LLMs overestimate sustainable engagement, understate the intention-behavior gap, disproportionately inflate engagement in low-impact behaviors relative to high-impact behaviors, and, more importantly, amplify demographic stereotypes about who is perceived as sustainable. In Study 2, we examine LLMs in their prescriptive role by testing whether ChatGPT-inferred demographic groups influence recommended sustainability behaviors and how personalization capability shapes these recommendations in a sample of about 4,000 AI users stratified across the same five demographic dimensions and split between two countries that vary in personalization capability. We find systematic differences in recommendations: inferred demographic group membership predicts which sustainability behaviors are recommended, with higher-impact behaviors directed toward groups that, on average, bear lower responsibility for emissions. Greater personalization capability attenuates these differences, suggesting that recommendations rely less on inferred demographic group membership and more on individual-level information. These findings highlight the need to align LLM design and privacy regulations to support effective and equitable climate action.