The Basic B*** Effect: The Use of LLM-based Agents Reduces the Distinctiveness and Diversity of People’s Choices
Citation. Matz, S. C., Horton, C. B., & Goethals, S. (2025). The Basic B** Effect: The Use of LLM-based Agents Reduces the Distinctiveness and Diversity of People’s Choices*. arXiv. https://arxiv.org/abs/2509.02910
Summary. Matz, Horton, and Goethals test whether delegating identity-relevant choices to LLM-based agents changes the distinctiveness and diversity of people’s preferences. Using real Facebook page-following choices from 1,000 U.S. users, they compare a human baseline against generic and personalized GPT-4o agents choosing between options the user had actually followed. Both agent types shift choices toward more popular options, while personalization creates a trade-off: it buffers between-person homogenization but more strongly compresses within-person choice diversity. The paper is useful for the wiki because it treats agentic AI not only as task automation, but as a decision architecture that can reshape agency, identity, and preference variety.
Key claims
- Delegating binary preference choices to LLM-based agents reduced interpersonal distinctiveness relative to the human baseline: generic agents selected less unique pages, t(999) = -11.02, p < .001, d = 0.21, and personalized agents also reduced distinctiveness, t(999) = -4.08, p < .001, d = 0.08.
- The generic agent’s distinctiveness loss was more than 2.5 times larger than the personalized agent’s, t(999) = 7.10, p < .001, d = 0.13, suggesting personalization can partially preserve uniqueness across people.
- Personalized agents produced a stronger compression of intrapersonal choice diversity: topical diversity fell, t(999) = -16.71, p < .001, d = -0.31, and psychological diversity fell, t(999) = -5.69, p < .001, d = -0.09.
- Generic agents also reduced topical diversity, t(999) = -10.80, p < .001, d = -0.21, but increased psychological interest diversity relative to baseline, t(999) = 7.62, p < .001, d = 0.12, indicating that generic and personalized agents create different distinctiveness-diversity trade-offs.
- The mechanism proposed is not simple error or low-quality advice: because both options came from the user’s own revealed preferences, the risk is that agentic optimization pulls people toward statistically likely or personally typical choices while narrowing exploration.
Evidence & limitations
- The study uses 110,000 choices from 1,000 U.S. users in the myPersonality Facebook dataset; Facebook pages are treated as revealed-preference proxies, which gives behavioral grounding but limits generalizability to other choice domains such as shopping, music, news, work, or social interaction.
- The AI conditions used GPT-4o as a generic recommender or as a personalized agent with age, gender, ten followed pages, and a generated profile summary from status updates; the design tests AI choice behavior over known user-liked alternatives rather than whether users would accept AI-generated recommendations in live settings.
- The Drive PDF hash matches the public arXiv PDF exactly; Crossref exact-title lookup did not return a matching published DOI as of ingest.
- The source is an arXiv preprint, so the evidence should be treated as public but not yet peer-reviewed.