DisprovedChanged2022

A computer cannot create an original image from a written description

The old claim turned out to be wrong and was rejected by evidence.

TaughtSoftware can edit, filter, and rearrange existing pictures, but it cannot invent a new image from scratch. Describe 'an astronaut riding a horse in the style of a Renaissance painting' and no program could produce it; visual creativity is human.

NowIn 2022, text-to-image models made this routine. OpenAI released DALL·E 2 in April, and Stable Diffusion was released openly in August: systems that generate novel, coherent, high-resolution images from a plain-English prompt in seconds.

What actually happened

These models learned the statistical relationship between text and pictures from enormous image–caption datasets, then generate by starting from noise and 'denoising' it toward an image that matches the prompt. The results were not collages of existing pictures but genuinely new compositions, including combinations, that astronaut on horseback, no one had photographed.

2022 was the hinge year: DALL·E 2 behind a waitlist in spring, Stable Diffusion's open release in late summer putting the capability on ordinary laptops. The honest caveats are real and unresolved: the training data, the effect on working artists, the model's habit of confidently mangling hands and text. But the flat claim that a computer cannot originate an image from a description stopped being true.

Newsletter