Intellectual property comprises legal rights protecting creations of the mind, including artistic works and inventions. In text-to-image generation, it centers on whether training data and generated images copy protected works or transform learned patterns into new expression. section 13.2 section 1.5.1
The debate divides into two positions:
- Critics argue that models may ingest copyrighted artworks without permission and reproduce distinctive styles, compositions, elements, or nearly identical images. section 1.5.1
- Supporters argue that models learn statistical relationships among pixels, objects, and styles, then combine concepts into new outputs rather than directly memorizing individual works. section 1.5.1
- Fair use may support transformative training, but the scale of copyrighted data and direct ingestion without permission can also be viewed as infringement. section 13.2
Near-identical generation
A prompt closely related to a known artwork may produce an image that appears nearly identical, illustrating why the boundary between generation and copying remains contested. section 1.5.1
