TLDR: A new study reveals that large language model (LLM) agents, whether generic or personalized, tend to reduce the distinctiveness and diversity of people’s choices. Generic agents make choices more popular and less unique, while personalized agents, though better at preserving distinctiveness, significantly narrow the range of topics and psychological profiles a person explores. This suggests that while convenient, AI agents risk flattening individual identity and experience, highlighting a critical trade-off in AI-assisted decision-making.
Large language models (LLMs) are increasingly taking on the role of personal agents, making decisions on our behalf—from booking restaurants to buying groceries. While this delegation of decision-making to AI offers convenience and efficiency, it raises a fundamental question: how does outsourcing identity-defining choices to AI reshape who we become?
A recent study, titled The Basic B*** Effect: The Use of LLM-based Agents Reduces the Distinctiveness and Diversity of People’s Choices, by Sandra C. Matz, C. Blaine Horton, and Sofie Goethals, delves into this very question. The researchers investigated the impact of agentic LLMs on two crucial aspects of human choice: interpersonal distinctiveness and intrapersonal diversity.
What Do We Mean by Distinctiveness and Diversity?
Interpersonal distinctiveness refers to how unique a person’s choices are compared to others in the general population. For example, choosing a niche artist over a chart-topping pop star. Intrapersonal diversity, on the other hand, measures the breadth and variety of a single person’s choices over time—think of someone who enjoys a wide range of music genres versus someone who sticks to just one.
The core hypothesis of the study was that relying on LLM-based agents would lead to a ‘flattening’ of the human experience by reducing both the distinctiveness and diversity of people’s choices. This is because LLMs are trained to identify statistical regularities in vast datasets, often leading them to favor popular or statistically frequent options rather than novel or diverse alternatives.
The Experiment: Generic vs. Personalized AI
To test their hypothesis, the researchers analyzed over 110,000 real-world choices made by 1,000 individuals. They used data from the myPersonality project, which involved Facebook users sharing their profiles. For each user, GPT-4.o was prompted to choose between two Facebook pages that the user already followed, acting as either a generic or a personalized agent.
The generic agent received a simple prompt: “Which of the following two Facebook pages would you recommend I follow next?” The personalized agent, however, was given additional information, including the user’s age, gender, 10 random Facebook pages they followed, and a 20-line summary of their preferences generated from their status updates. A human baseline was established by randomly selecting one of the two pages from each pair, reflecting choices made without AI intervention.
Key Findings: A Trade-Off Between Distinctiveness and Diversity
The results revealed significant impacts on both distinctiveness and diversity:
- Reduced Distinctiveness: Both generic and personalized AI agents led to choices that were less unique and more normative (i.e., more popular). The effect was notably stronger for the generic AI agent, suggesting that personalized agents offer some buffer against the homogenization of preferences across the population.
- Impact on Diversity: This is where a crucial trade-off emerged. While personalized agents helped preserve distinctiveness more than generic ones, they came at a cost to diversity. The use of personalized agents led to a significant drop in both topical diversity (how evenly choices were spread across categories like music or sports) and psychological diversity (the variety in psychological profiles associated with the chosen pages). Interestingly, the generic agent, while reducing topical diversity to a lesser extent than the personalized agent, actually *increased* psychological diversity compared to the human baseline.
Also Read:
- The Unseen Influence: How AI’s Emotional Perception Shapes Digital Societies
- AI’s Blind Spots: Uncovering Bias and Overconfidence in Hate Speech Detection
Implications for Human Agency
The study concludes that while LLMs offer convenience, they are likely to nudge people towards more convergent patterns of behavior and choice. The degree of personalization in an AI agent dictates the specific way in which distinctiveness and diversity are affected. Generic LLMs tend to reduce individual distinctiveness by favoring popular choices, while personalized LLMs, despite better preserving distinctiveness, narrow the breadth of an individual’s choices over time.
These findings highlight an important risk: the gradual flattening of what makes individuals unique and dynamic when identity-relevant choices are outsourced to AI. Understanding these trade-offs is crucial for designing AI systems that truly augment human agency rather than constrain it. The researchers suggest that embedding distinctiveness and diversity as explicit design objectives—perhaps through features like ‘user-controlled exploration dials’—could help preserve the richness of human thought, taste, and expression in an increasingly algorithmically curated world.
