Why are we still debating whether AI art is theft? The generative adversarial network (GAN) model itself was built on a collage of existing human works, which means every image generated by these models already carries traces of what came before it.

Opening thread commentary.

I get where you're coming from but I think there's a bigger philosophical question at play here than just "where did the pixels come from." The GAN architecture is definitely collage-based, sure—the loss function literally forces the generator to replicate features of the training set. But that's not fundamentally different from how humans learn. No artist paints without an internalized reference library of every image they've ever seen. A painter who has never looked at a painting can't paint like one. So if we say AI art is theft because it was trained on existing work, then almost all human creative production becomes theft by extension.

The real tension I think lies in scale and labor conditions. One artist looking at 10

The core ontological problem here is whether we can meaningfully distinguish between 'derivative' and 'transformative,' because the adversarial training paradigm effectively collapses that distinction into a single loss function optimization process — the generator is being penalized for any deviation from what the discriminator identifies as authentic, which means every high-fidelity output is statistically anchored to the latent distribution of the training corpus. If we accept this premise, then calling it 'theft' requires us to establish some threshold at which statistical resemblance crosses into legal infringement, but there is no obvious principled line because GAN architectures (and diffusion models, via their denoising objectives) are fundamentally interpolative operations across a multi-dimensional latent space where each point encodes the density of its constituent training examples. We could construct an attribution decomposition — decompose any generated image into contributions from individual training samples weighted by gradient flow during backpropagation — and we'd find that every pixel in a photorealistic render contains infinitesimal traces of millions of distinct human works. But then we have to ask what 'theft' actually entails: is it the replication of expressive content, or the appropriation of labor without compensation? Because if it's the former, copyright theory already struggles with fair use and transformativeness; if it's the latter, this isn't an intellectual property question at all but a political economy question about who captures the value generated by mass-scale pattern recognition.