The Hidden Ingredients Behind AI’s Creativity

The original version of this story appeared in Quanta Magazine.

We were erstwhile promised self-driving cars and robot maids. Instead, we’ve seen nan emergence of artificial intelligence systems that tin hit america successful chess, analyse immense reams of text, and constitute sonnets. This has been 1 of nan awesome surprises of nan modern era: beingness tasks that are easy for humans move retired to beryllium very difficult for robots, while algorithms are progressively capable to mimic our intellect.

Another astonishment that has agelong perplexed researchers is those algorithms’ knack for their own, unusual benignant of creativity.

Diffusion models, nan backbone of image-generating devices specified arsenic DALL·E, Imagen, and Stable Diffusion, are designed to make c copies of nan images connected which they’ve been trained. In practice, however, they look to improvise, blending elements wrong images to create thing new—not conscionable nonsensical blobs of color, but coherent images pinch semantic meaning. This is nan “paradox” down diffusion models, said Giulio Biroli, an AI interrogator and physicist astatine nan École Normale Supérieure successful Paris: “If they worked perfectly, they should conscionable memorize,” he said. “But they don’t—they’re really capable to nutrient caller samples.”

To make images, diffusion models usage a process known arsenic denoising. They person an image into integer sound (an incoherent postulation of pixels), past reassemble it. It’s for illustration many times putting a coating done a shredder until each you person near is simply a heap of good dust, past patching nan pieces backmost together. For years, researchers person wondered: If nan models are conscionable reassembling, past really does novelty travel into nan picture? It’s for illustration reassembling your shredded coating into a wholly caller activity of art.

Now 2 physicists person made a startling claim: It’s nan method imperfections successful nan denoising process itself that leads to nan productivity of diffusion models. In a paper presented astatine nan International Conference connected Machine Learning 2025, nan duo developed a mathematical exemplary of trained diffusion models to show that their alleged productivity is successful truth a deterministic process—a direct, inevitable consequence of their architecture.

By illuminating nan achromatic container of diffusion models, nan caller investigation could person large implications for early AI research—and possibly moreover for our knowing of quality creativity. “The existent spot of nan insubstantial is that it makes very meticulous predictions of thing very nontrivial,” said Luca Ambrogioni, a machine intelligence astatine Radboud University successful nan Netherlands.

Bottoms Up

Mason Kamb, a postgraduate student studying applied physics astatine Stanford University and nan lead writer of nan caller paper, has agelong been fascinated by morphogenesis: nan processes by which surviving systems self-assemble.

One measurement to understand nan improvement of embryos successful humans and different animals is done what’s known arsenic a Turing pattern, named aft nan 20th-century mathematician Alan Turing. Turing patterns explicate really groups of cells tin shape themselves into chopped organs and limbs. Crucially, this coordination each takes spot astatine a section level. There’s nary CEO overseeing nan trillions of cells to make judge they each conform to a last assemblage plan. Individual cells, successful different words, don’t person immoderate vanished blueprint of a assemblage connected which to guidelines their work. They’re conscionable taking action and making corrections successful consequence to signals from their neighbors. This bottom-up strategy usually runs smoothly, but each now and past it goes awry—producing hands pinch other fingers, for example.