Serendipity by Design: Evaluating the Impact of Cross-domain Mappings on Human and LLM Creativity

Who Should Read

Researchers studying AI creativity and product teams using LLMs for ideation who want to understand the best prompting strategies for creative tasks.

Core Mechanics

• LLMs outperform average humans on divergent creativity benchmarks (Alternative Uses Task, Unusual Uses) when using standard prompting
• Human creativity techniques like analogical reasoning ('think like someone from field X') significantly boost human performance but provide little or no benefit for LLMs
• LLMs' creative strength is breadth — they generate many diverse ideas. Their weakness is depth — breakthrough or truly novel ideas remain rare
• Prompting LLMs to use human creativity techniques (SCAMPER, random word association, etc.) shows inconsistent results — sometimes helping, often neutral or harmful
• The 'inspiration from other fields' technique works for humans because it forces perspective shifts they wouldn't naturally take — LLMs already have broad cross-domain knowledge and don't gain from forced perspective shifts
• Best prompting strategy for LLM ideation: high temperature, explicit diversity instructions, and iterative refinement outperform creative technique instructions

Evidence

• On Alternative Uses Task: LLM average originality score 3.8/5 vs. human average 2.9/5 under standard conditions
• Cross-domain inspiration prompt for humans: +0.7 originality points. For LLMs: +0.05 (statistically insignificant)
• SCAMPER technique: improved human ideation by 23%, improved LLM ideation by 2% (within noise)

How to Apply

• For LLM ideation: skip the human creativity technique instructions. Instead, use: high temperature (0.9+), explicit 'be unconventional and avoid obvious answers' instruction, and ask for 20+ ideas then filter.
• Don't copy human workshop facilitation techniques directly into LLM prompts — they're designed to overcome human cognitive biases that LLMs don't have.
• Use LLMs for breadth (generating many diverse starting points) and humans for depth (selecting and developing truly novel directions) — play to each's strengths.

Code Example

# Cross-domain mapping prompt example (forcing semantically distant source assignment)
import random

products = ["backpack", "smartphone", "sofa", "chef's knife", "sneakers"]
sources = ["octopus", "tornado formation", "coral reef", "symphony orchestra", "hydroelectric dam", "GPS navigation"]

target = random.choice(products)
source = random.choice(sources)

prompt = f"""
You are a product designer. Your task is to invent a novel feature for a {target}.

Step 1: Identify a specific, interesting property or mechanism of a {source}.
Step 2: Map that property onto the {target} — how could this mechanism be translated into a real product feature?
Step 3: Describe the resulting product feature in 1-2 sentences, focusing only on what the product does (not the analogy itself).

Target product: {target}
Inspiration source: {source}
"""

print(prompt)
# Example output:
# Target product: backpack
# Inspiration source: hydroelectric dam
# → "This backpack converts kinetic energy from your movement into stored power in a built-in battery, letting you charge devices on the go."

Terminology

Related Resources

• Open Science Framework Pre-registration (Evaluation Experiments)