---
id: "concept-cosmetic-ai-diversity"
type: "concept"
source_timestamps: ["§ Agentic AI's Diversity Challenge"]
tags: ["prompt-engineering", "ai-personas", "illusion-of-diversity"]
related: ["concept-structural-ai-diversity", "contrarian-costume-change", "quote-costume-change", "entity-enver-cetin"]
definition: "The superficial practice of prompting a single foundation model to adopt different personas or cultural attitudes, failing to achieve true cognitive diversity."
sources: ["agentic"]
sourceVaultSlug: "hbr-seg-agentic"
originDay: 6
articleStem: "hbr-new-28-agent-teams-different-models"
sourceUrl: "https://hbr.org/2026/06/the-strongest-teams-of-ai-agents-will-be-built-using-different-models"
sourceTitle: "The Strongest Teams of AI Agents Will Be Built Using Different Models"
---
# Cosmetic AI Diversity

Cosmetic AI diversity occurs when organizations attempt to create varied AI agents by simply prompting a *single underlying foundation model* to adopt different personas, personality types (e.g., 'hot' vs. 'cold', extrovert vs. introvert), or cultural attitudes. While this creates surface-level variation, it fails to deliver true cognitive diversity because the underlying 'brain' (the foundation model), the retrieval architectures, and the data sources remain **identical**.

Research shows that prompting for personality types often produces highly **binary, extreme behaviors** rather than the nuanced continuum seen in humans; Big Five profiles elicited via prompting tend to be relatively stable and to yield exaggerated, non-human distributions. As [[entity-enver-cetin]] puts it (see [[quote-costume-change]]): **"Costume change is not cognition."**

This concept is the foil to [[concept-structural-ai-diversity]] and the crux of the article's core contrarian claim (see [[contrarian-costume-change]]).

**Enrichment nuance:** Industry guidance (IBM, AWS, Stanford HAI) likewise treats prompt-level variation as *configuration*, not a change to the underlying cognitive architecture. A fair counterpoint, however, is that persona prompts can still elicit *functionally useful* variance (different trade-offs, different parts of a model's knowledge) for brainstorming or scenario planning — useful, even if not structural.