Serendipity by Design: When Humans and AI "Brainstorm" Together

> Paper: Serendipity by Design: Evaluating Cross-domain Mappings on Human and LLM Creativity > Authors: Qiawen Ella Liu, Marina Dubova, et al. > arXiv: 2603.19087

Opening: An Umbrella Inspired by an Octopus

Imagine you are a product designer tasked with improving an ordinary umbrella. Without special inspiration, you might suggest: a new color, some patterns, sturdier ribs—all "obvious" improvements.

But what if you're told: draw inspiration from an octopus to improve the umbrella?

Your brain might stall for a moment, then start racing:

• An octopus has eight arms sensing water flow in all directions... could an umbrella sense wind direction all around?
• Octopuses squirt ink to escape... could an umbrella quickly collapse or release some signal?
• Octopus suckers adhere to surfaces... could an umbrella stick to a wall?

Suddenly your thinking opens up. This technique—cross-domain mapping, forcibly connecting seemingly unrelated things—is a fascinating phenomenon in creativity research. But the question is: does it work equally well for humans and AI?

That's what a Princeton research team set out to answer.

Background: The Mystery of "Remote Association"

What Is Creativity?

A classic theory comes from psychologist Sarnoff Mednick's 1962 Associative Hierarchy model. Think of your brain as a giant network where concepts are nodes:

• "Apple" → "fruit": near association
• "Apple" → "gravity" (Newton's story): medium-distance association
• "Apple" → "IPO": remote association

Mednick argued that highly creative people have a "flat" associative hierarchy—they can freely jump to distant concept nodes. This is why many great inventions came from cross-domain inspiration:

• The Wright brothers studied bird flight to design airplanes
• George de Mestral observed burdock burrs to invent Velcro
• Da Vinci studied water flow to design bridges and flying machines

Design Fixation: Why We Need "External Force"

Humans have a natural cognitive flaw—design fixation. Asked to design a new coffee cup, your first thoughts are anchored to existing features: bigger? a lid? a handle? Cross-domain mapping forces your brain out of its comfort zone, building connections you wouldn't normally make.

Experiment Design: Humans vs. AI

Research question: Do cross-domain mapping prompts equally boost human and LLM creativity?

Setup:

• Human participants recruited via crowdsourcing; AI participants included mainstream LLMs (e.g., GPT-4)
• Task: design new features for 10 everyday products (backpack, TV, bicycle, alarm clock, coffee cup, umbrella, desk, car, watch, refrigerator)
• Two prompt conditions:
1. User-need condition: "Design a new backpack feature addressing unmet user needs" 2. Cross-domain mapping condition: "Draw inspiration from 'cactus' to design a new backpack feature"—with randomly assigned source domains (octopus, cactus, GPS, bees, spider webs, sunflowers, bats, coral reefs, dandelions, camels, seashells, etc.)

Evaluation: Human raters scored ideas on originality, feasibility, and usefulness. Researchers also computed semantic distance between source and target domains using word vectors trained on Wikipedia.

Results: A Surprising Asymmetry

Finding 1: LLMs Are Naturally More "Out There"

Without cross-domain prompts, LLM-generated ideas were more original on average than human ideas. LLMs may already possess a "flat associative hierarchy"—their training data spans massive cross-disciplinary connections. Humans, by contrast, easily fall into design fixation.

Finding 2: Cross-Domain Mapping Helps Humans, But Barely Affects LLMs

The core finding:

• For humans: cross-domain prompts significantly improved originality
• For LLMs: no significant effect—scores were basically flat

Possible explanations: 1. LLMs are already "flat": with such broad training data, explicit prompting just executes an already-natural operation. 2. Humans need a "nudge": cross-domain mapping acts as a forced boost out of mental anchoring. 3. Distance-dependent prompt effects: the effect grows with semantic distance; for LLMs, explicit prompts may only add value at extreme distances (e.g., "cars and octopuses").

Finding 3: Semantic Distance Drives Originality

For both humans and LLMs: the greater the semantic distance between source and target domains, the more original the ideas. Inspiration from "birds" for "airplanes" yields conservative ideas; from "octopuses" for "umbrellas," wilder ones. This supports the long-standing hypothesis that remote association is a core mechanism of creativity.

Finding 4: Different Creativity Distributions

• Humans: a "bimodal" distribution—many mediocre ideas plus a few brilliantly original strokes
• LLMs: more uniform—higher overall originality, but lacking extreme outliers

Human creativity is sparse but explosive; LLM creativity is stable but capped.

Implications

For Human Creativity Training

Cross-domain mapping genuinely works, especially in early design stages when you feel stuck:

• Designers can use "random word" techniques in brainstorming
• Teams can hold regular "cross-domain learning" sessions—programmers attend concerts, designers read scientific papers
• Companies can build "analogy databases" of cross-industry innovation cases

For AI-Assisted Design

1. Don't expect explicit prompts to "inspire" LLMs: if you're already using AI for ideation, repeatedly asking it to "be wilder" may be unnecessary—it may already be there. 2. New human+AI collaboration model: humans handle "cross-domain leaps"; AI handles "idea expansion" from human inspiration seeds. 3. Semantic distance as a quality predictor: AI tools could compute the semantic distance between recommended inspiration sources and the target problem, prioritizing mappings that are "far enough but not absurd."

For Understanding Creativity Itself

• Creativity is multidimensional: originality, feasibility, and usefulness often trade off.
• Randomness catalyzes creativity: randomly connecting unrelated things explains why many breakthroughs come from serendipity.
• Human and AI creativity are complementary: humans excel at extreme leaps; AI at consistent quality. Combined: 1+1>2.

Conclusion: The Designed Art of Chance

The title "Serendipity by Design" combines two seemingly contradictory ideas:

• Serendipity: randomness, unpredictability
• Design: purpose, controllability

The research shows creativity can be partly "designed"—by systematically introducing cross-domain mappings, we can manufacture the happy accidents that otherwise depend on luck. But it also reveals a boundary: this "designed serendipity" works differently for humans and AI. Humans need the nudge to escape fixation; AI may already live in a permanently "flat associative space."

Perhaps the best creativity system is neither pure AI nor pure human, but a collaboration where humans "open the brainstorm" and AI "fills in the details"—like that octopus-inspired umbrella: the human supplies the "all-around sensing" inspiration, and the AI helps design the sensor layout and algorithms. That is the future of human-AI co-creativity.

Further reading:

• Paper: https://arxiv.org/abs/2603.19087
• Related concepts: remote association theory, structure mapping theory, semantic distance, design fixation
• Classic reading: Mednick (1962), "The Associative Basis of the Creative Process"