The hidden ingredients behind AI’s creativity

undefined 26, 2025

Source: wired

Researchers have uncovered why diffusion-based AI models appear creative: their so-called creativity arises from technical quirks in the denoising process, locality and equivariance, which force models to generate image patches independently and then assemble them. This mechanism, tested with the ELS machine, can predict AI outputs with ~90% accuracy, suggesting AI creativity may be a predictable by-product of its architecture, offering insights into both machine and human creativity.

Share on