The Language of Creativity: Evidence from Humans and Large Language Models

Area of Science:

• Computational linguistics
• Psychology of creativity
• Artificial intelligence

Background:

• Automated assessments of verbal creativity are advancing through computational scoring and semantic distance.
• Previous research focused on semantic diversity as a predictor of creative text.

Purpose of the Study:

• To extend computational linguistic approaches for characterizing creative text features.
• To investigate the predictive power of perceptual details and semantic diversity on creativity.
• To explore generative language models (like GPT-3 and GPT-4) for mimicking human creativity and supplementing data collection.

Main Methods:

• Collected 600 short stories from human participants and GPT-3.
• Assessed creative quality of randomized stories.
• Replicated findings with 120 stories generated by GPT-4.
• Analyzed stories for semantic diversity and presence of perceptual details.

Main Results:

• Perceptual details, alongside semantic diversity, strongly predict creativity.
• No significant difference in creativity ratings between human and AI-generated stories.
• Positive correlations observed between human and AI assessments of creativity.

Conclusions:

• Perceptual details are crucial for creative writing, enhancing reader transportation.
• Generative language models demonstrate human-like creative writing capabilities.
• Computational linguistic methods offer robust tools for creativity assessment.