---
id: "lib-2026-10-12-creative-destruction-human-ai-competition-in-an"
type: "library"
title: "Creative destruction: Human-AI competition in an experimental market"
lang: "EN"
org: "arXiv (Ohio State, CMU)"
source_type: "Paper"
date: "2026-10-12"
week: "2026-W42"
url: "https://nosov.ai/library/lib-2026-10-12-creative-destruction-human-ai-competition-in-an/"
original_url: "https://arxiv.org/abs/2610.18921"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Preregistered, incentivised experiment by economists at Ohio State and Carnegie Mellon with N = 1,025 and over 1,600 real games; clear design and effect sizes, though the market is a lab setting."
summary_basis: "full-text"
original_minutes: "30"
tags: ["work and ai", "creative work", "consumer demand", "experiments"]
---

# Creative destruction: Human-AI competition in an experimental market

Tier A (Strong evidence): Preregistered, incentivised experiment by economists at Ohio State and Carnegie Mellon with N = 1,025 and over 1,600 real games; clear design and effect sizes, though the market is a lab setting.

Summary based on the full text.

## In short

John Conlon (Ohio State) and Peter Schwardmann (Carnegie Mellon) test two common reasons people give for why humans will stay in production as AI improves: humans have better taste, or buyers prefer human-made goods. In three preregistered studies, participants directed an AI coding agent to improve simple games, and other participants paid real money to play them. Games designed by a single AI prompt beat games directed by amateurs and by professional creatives. Telling buyers who made the game did not change what they paid.

## Key numbers

- $0.47 vs $0.37: average willingness to pay for AI-designed vs human-designed games (gap 0.31 SD, p < 0.001)
- 0.16 SD: how far professional creatives' games trailed AI-only games on willingness to pay
- 59.3% vs 58.2%: probability blind consumers assigned to a game being human-made, for AI vs human games
- 6.8 vs 6.0 cents: AI premium in the blind vs disclosed arms; no significant difference (p = 0.79)

## What it reports

- Consumers paid $0.37 on average for a human-directed game over its basic version, and $0.47 for AI-designed games; a 10-cent gap, 0.31 SD.
- Professionals (game developers, graphic designers, content creators) trailed AI games by 0.16 SD. On a combined quality index, professionals scored 0.21 SD and amateurs 0.34 SD below AI.
- Consumers rated AI games more fun, more creative and more polished. AI prompts were longer (235 words vs 89 for professionals and 60 for amateurs), suggesting more ideas per prompt.
- Producers expected to beat the AI, by 9 cents (amateurs) and 17 cents (professionals), until they read the AI's idea.
- Disclosure did not move demand: AI premium 6.8 cents blind, 6.0 cents disclosed. Blind consumers could not tell AI from human games.
- Humans who watched the AI design a game improved, but still trailed it by 0.21 SD (Study 1) and 0.15 SD (Study 3).
- 80.7% of producers were at least moderately proud of their game.

## How they got it

1,025 US participants on Prolific: 296 as producers and consumers, 180 professionals as producers only, 549 as consumers only. Producers improved basic games over three rounds of messages to an Opus 4.8 coding agent. AI-only games came from one Sonnet 4.6 prompt passed to the same agent, with no testing or iteration. Willingness to pay was elicited with an incentive-compatible rule and paid to producers.

## Limits

- Short, simple browser games played for about a minute; results may not carry to complex products or long projects.
- The AI was deliberately handicapped (one prompt, no testing), which strengthens the finding but limits what it says about human-AI teams with more time.
- Online participants; professionals were identified by platform screeners and self-report.

## Why it matters for companies

The finding cuts against the comfortable claim that human taste will direct AI for the foreseeable future, at least for small creative tasks. Firms should test AI-only baselines before assuming a human-in-the-loop design adds value, and measure with real customer choices. The overconfidence result matters for managers: staff may not see the gap until shown the AI's output. Disclosure labels may not protect human-made products with buyers.

Original: https://arxiv.org/abs/2610.18921