Introduction
As generative AI becomes ubiquitous, a central question persists: Are these models truly creative, or are they merely "stochastic parrots" rearranging training data? This survey addresses the lack of a unified framework for evaluating machine creativity, distinguishing between high-quality interpolation (averaging known patterns) and genuine invention. By analyzing the current state of the art, we identify where AI has made genuine progress and where it hits a "creativity ceiling"—particularly in tasks requiring abstract reasoning, compositionality, and intentionality.
A Taxonomy of AI Creativity
To evaluate progress, we must define the dimensions of creativity. We synthesize research through several psychological and computational lenses: the Four-C Model (categorizing creativity from personal "mini-c" learning to world-changing "Big-C" genius), Boden’s Three Types (combinatorial, exploratory, and transformational), and a focus on Process vs. Product. Evaluating the internal logic and response pathways of a model is as critical as assessing its final output.
Domains and Benchmarks
AI creativity is tested across linguistic, artistic, problem-solving, and scientific domains. Key benchmarks include the Alternative Uses Task (AUT) for divergent thinking, the Abstraction and Reasoning Corpus (ARC) for fluid intelligence, and specialized datasets for humor, metaphor, and scientific discovery (e.g., AI Feynman for equation discovery).
Key Insights: The Creativity Ceiling
Current AI excels at interpolation—filling gaps within known data distributions. However, it struggles with extrapolation and invention:
- The "Einstellung" Effect: Models often exhibit functional fixedness, biased toward common associations (e.g., failing to see a shoe as a potential hammer because its training insists a shoe is for walking).
- The "Salieri" Effect: LLM outputs tend to be more homogenous and less exceptional than human-written text. They are proficient at the "average" but struggle with the "exceptional."
- Lack of Agency: AI lacks spontaneity and intentionality, creating only when prompted and following statistical probability rather than intentional divergence.
Empirical Results
- The ARC Gap: On the Abstraction and Reasoning Corpus (ARC), the best AI models currently score around 49.5%, while the human-level threshold is 85%.
- Idea Uniqueness: In large-scale studies of AI idea generation, only a tiny fraction of generated ideas (e.g., 5%) were found to be truly unique or original.
- Compositional Failure: Models frequently fail at "negation" and "spatial relationships" in image generation (e.g., rendering "a bowl with fewer strawberries than cucumbers").
Future Directions
Overcoming the "creativity ceiling" requires addressing several grand challenges:
- Active Divergence: Designing architectures that intentionally move away from the most likely training distribution to find novel solutions.
- Semantic Planning: Moving beyond token-by-token prediction toward high-level planning to maintain long-range coherence in stories and scientific theories.
- Human-AI Collaboration: Developing "centaur" systems where AI acts as a creative assistant to human "Big-C" creators.
- Ethics and Copyright: Navigating the legal tensions of training on copyrighted works to produce "new" art.
Summary updated