---
id: "contrarian-costume-change"
type: "contrarian-insight"
source_timestamps: ["§ Agentic AI's Diversity Challenge"]
tags: ["prompt-engineering", "ai-myths"]
related: ["concept-cosmetic-ai-diversity", "quote-costume-change", "concept-structural-ai-diversity"]
challenges: "The belief that prompt engineering alone is sufficient to generate diverse perspectives and avoid groupthink in AI systems."
speakers: ["Enver Cetin", "Mark Purdy"]
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"
---
# Prompting for personas does not create diverse AI

**Contrarian insight.**

**Conventional wisdom:** You can create a diverse team of AI agents by instructing a single LLM to act like different people (e.g., 'You are a skeptical risk manager' vs. 'You are an optimistic marketer').

**The inversion:** This is merely a **'costume change'** (see [[quote-costume-change]]). Because the underlying foundation model, training data, and retrieval architecture are **identical**, the agents lack true cognitive diversity and still suffer from [[concept-correlated-ai-errors]]. This is the core of [[concept-cosmetic-ai-diversity]]; the real fix is [[concept-structural-ai-diversity]].

**Challenges:** The belief that prompt engineering alone is sufficient to generate diverse perspectives and avoid groupthink in AI systems.

**Enrichment — steel-man the other side:** Persona prompting is not *worthless*. Even without changing cognition, it can (1) elicit different parts of a model's knowledge, (2) encourage different trade-offs (risk-averse vs. risk-seeking), and (3) offer useful variance for brainstorming or scenario planning — and evaluation frameworks routinely use LLM-as-judge with varied prompts to simulate diverse evaluators. So the sharper framing is: persona prompting provides *functional* variance but not *structural* diversity; dismissing it as purely cosmetic may overlook genuine utility.