This site is a work in progress and has not been widely shared. Content may contain errors. Feedback is welcome.
This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.
Back to papers

Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content

Doshi, Hauser

2024Science Advances350 citations
Experimental evidenceInterdisciplinaryCausal
LLM / Generative AIWriting / contentCreative workHuman-AI collaborationAugmentation vs. substitution
DOI: 10.1126/sciadv.adn5290
Abstract

Creativity is core to being human. Generative artificial intelligence (AI)—including powerful large language models (LLMs)—holds promise for humans to be more creative by offering new ideas, or less creative by anchoring on generative AI ideas. We study the causal impact of generative AI ideas on the production of short stories in an online experiment where some writers obtained story ideas from an LLM. We find that access to generative AI ideas causes stories to be evaluated as more creative, better written, and more enjoyable, especially among less creative writers. However, generative AI–enabled stories are more similar to each other than stories by humans alone. These results point to an increase in individual creativity at the risk of losing collective novelty. This dynamic resembles a social dilemma: With generative AI, writers are individually better off, but collectively a narrower scope of novel content is produced. Our results have implications for researchers, policy-makers, and practitioners interested in bolstering creativity.

Summary

Doshi and Hauser conduct a pre-registered two-phase online experiment with 293 writers randomly assigned to write stories with or without access to GPT-4-generated ideas, evaluated by 600 third-party raters, to study how generative AI affects human creativity in an open-ended creative task.

Main Finding

Access to GenAI ideas significantly increases individual creativity (novelty +6.7%, usefulness +6.4%) and story quality, with largest gains for less creative writers (10-11% improvement), but reduces collective diversity as GenAI-assisted stories become 5% more similar to each other than human-only stories.

Primary Datasets

Doshi-Hauser Creative Writing

AI-focused

Prolific online experiment (N=293 writers, N=600 evaluators)

Secondary Datasets

OpenAI GPT-4 API for generating story ideas and text embeddings; Divergent Association Task (DAT) scores

Key Methods
Two-phase randomized online experiment with 293 writers assigned to conditions (Human only, Human with 1 GenAI idea, Human with 5 GenAI ideas) and 600 third-party evaluators rating stories. OLS regression with robust standard errors, intention-to-treat analysis.
Sample Period
2023
Geographic Coverage
United Kingdom (Prolific participants)
Sample Size
293 writers producing stories; 600 evaluators providing 3,519 evaluations
Level of Analysis
Individual
Occupation Classification
None
Industry Classification
None
Notes
Science Advances, vol. 10, no. 28 [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions. [Claude classification]: Published in Science Advances vol. 10, no. 28 (2024). Pre-registered at AsPredicted.org (ID 136723). Uses cosine similarity of embeddings from OpenAI API to measure story similarity. Writers completed Divergent Association Task (DAT) to measure inherent creativity. The paper finds evidence of both enhancement (individual creativity improves) and homogenization (collective diversity decreases) effects. 88.4% of participants in GenAI conditions chose to access at least one AI idea. Evaluators were blind to treatment assignment initially, then informed for ownership/ethics questions.