The Struggle With AI Generated Art

In the span of about eighteen months, typing a sentence into a box and getting a finished illustration back went from a research demo to something anyone can run on a laptop or phone. Tools like OpenAI’s DALL-E and Stability AI’s Stable Diffusion take a prompt like “an astronaut eating cereal on Mars” and return four variations in seconds.

These systems use diffusion models, which train by adding Gaussian noise to millions of images and learning to reverse that process step by step, guided by text embeddings. You can adjust the prompt, inpaint a region, or feed an output back into the model until it matches what you want. As an ML engineer, I like the math behind latent diffusion. What it is doing to working illustrators and concept artists is much harder to defend.

Struggle From Artists

A human artist spends years developing a recognizable style through study, failed sketches, and practice. For working illustrators, that style is how clients recognize and hire them.

Image generators were trained on large web scrapes that included copyrighted portfolios without permission. Anyone can type “a person riding a horse by a river, in the style of Vincent van Gogh” (or in the style of a living freelance illustrator on ArtStation) and get a close imitation in five seconds.

For many commercial buyers, an AI image does not need to be great. It only needs to be good enough for a storyboard, a book cover, or a marketing banner at a fraction of the cost and turnaround time. Freelance illustration already had thin margins and heavy competition. Now illustrators are competing against models trained on their own portfolios without consent, credit, or pay.

Unlike traditional plagiarism, where a stolen image still carries a watermark or a file history, diffusion models blend their training data into billions of weights. An output can mimic a specific artist’s line work and palette without copying a single pixel directly from any one file.

Developer’s Perspective

Engineers and researchers building these models usually compare training to how humans learn. From their point of view, a diffusion model learning visual patterns from millions of images is similar to an art student studying paintings in a museum.

They also point to who uses these tools. Solo game developers, writers, and people who cannot draw can mock up scenes themselves, and working artists can use image models for quick rough drafts the same way programmers use autocomplete.

Where This Leaves Artists

Latent diffusion is impressive engineering, and at the same time the datasets behind it treated working artists as free training data. Without provenance tracking, opt-out enforcement, and licensing for training sets, the people whose work made these models possible get none of the upside.