AI Art Generators Learned by "Stealing" Art

When AI image generators became popular, a common claim emerged that these tools work by storing copies of copyrighted images and stitching them together. The narrative was that asking for "a painting in the style of [artist]" would retrieve and remix that artist's stolen work.

AI image generators do not store copies of their training images. A model like Stable Diffusion has about 2 GB of weights but was trained on billions of images. It is mathematically impossible to store those images in 2 GB. Instead, the model learns statistical patterns - textures, compositions, styles, concepts - and generates new images based on those patterns.

However, the ethical and legal concerns are real and nuanced. These models were trained on datasets that included copyrighted work scraped from the internet without explicit consent. While the models do not copy specific images, they learn from those images, and the artists were not compensated or asked for permission.

The "theft" framing is misleading because it implies direct copying, which is not how the technology works. But dismissing all concerns is equally wrong. There are legitimate questions about consent, compensation, and the economic impact on artists. Several lawsuits are working through these issues.

The honest framing: AI image generators learn from copyrighted work without permission (ethically concerning) but do not store or copy specific images (technically accurate). The legal and ethical frameworks are still being developed. Comparing AI learning from images to a human artist studying other artists' work is a common analogy, but it is imperfect because the scale and commercial application are fundamentally different.

Stability AI: How Stable Diffusion works - https://stability.ai/

EFF: AI and copyright - https://www.eff.org/

US Copyright Office: AI and copyright registration - https://www.copyright.gov/ai/

Vox: "The AI art controversy, explained" - https://www.vox.com/