Reproduce “Putting GPT-3's Creativity to the (Alternative Uses) Test” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Claire Stevenson, Iris Smal, Matthijs Baas, Raoul Grasman, Han van der Maas
A study from the University of Amsterdam assessed GPT-3's creative capabilities against human performance on Guilford's Alternative Uses Test, finding that humans produced more original and surprising ideas with greater semantic distance, whereas GPT-3 generated more useful responses.
Key insights
- Human creativity excels in divergent thinking (generating novel, unexpected ideas), while GPT-3 demonstrates strength in convergent thinking (finding practical, useful solutions).
- Creativity is multi-dimensional, meaning an AI system might perform well in certain aspects like utility but underperform in others such as originality and surprise.
- The findings suggest that the statistical pattern learning nature of LLMs may inherently limit their capacity to produce truly novel combinations that transcend their training data.
- Both humans and GPT-3 exhibit a negative correlation between originality and utility, indicating a fundamental trade-off between these creativity dimensions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.