Home / Technology / AI Image Creation Lacks Direct Artist Links
AI Image Creation Lacks Direct Artist Links
25 Aug
Summary
- AI models create images unlinked to specific training data.
- Removing art from training data doesn't alter AI output.
- Attribution decay makes IP theft claims difficult for artists.
A new study published in Nature Communications challenges the notion that AI image generators reference specific artworks. Researchers from MIT investigated diffusion models, the technology behind tools like Midjourney and Stable Diffusion, to understand their creative process. They discovered that these models' outputs are fundamentally "unattributable" to any single piece of training data.
The study introduced the concept of 'attribution decay,' demonstrating that as the training dataset size increases, it becomes impossible to link a generated image to its source material. Even if an AI-generated image closely resembles an artist's work, removing that specific work from the training data results in no discernible change to the AI's output.
This finding has significant implications for artists attempting to pursue legal action against AI companies for intellectual property infringement. The research suggests that proving direct copying is not feasible with current diffusion models. The authors highlight the 'alien' nature of these systems, which differ from human creativity where influences are often conscious.
