Who owns AI generated Artwork?
There has been much discussion lately regarding the copyright implications of AI-generated artwork programs. Two questions currently dominate the debate:
- Who owns the AI-generated output?
- What are the copyright implications of training inputs for these algorithms?
Clarification of the latter may be imminent with a class action lawsuit Andersen v. Stability AI Ltd. (3:23-cv-00201), District Court, N.D. California. filed in the US last week against Stability AI, Midjourney, and Deviant Art (‘’the Defendants”) for copyright infringement by artists whose works have been used as training images for the neural network.
Stable Diffusion is an image-generating engine and creates “new images” through a combination of conditioning data (text prompts) and latent images (a copy of an image from the training dataset). It is alleged that the images created by DreamStudio (by Stability), DreamUp (by Deviant Art) and the Midjourney Product, all based on Stable Diffusion, are generated exclusively from copies of copyrighted images. The Defendants are not licensed by the Plaintiffs to download, store, or distribute copies of the copyrighted works for use in training or creating derivative works or for their subsequent distribution.
The claims for relief include:
- Copyright infringement (direct and vicarious);
- Digital Millennium Copyright Act violations (including the removal of copyright management information from the copyrighted images);
- Rights of publicity violations (in designing the AI Image products allowing for output ‘in the style’ of specific images); and
- Unlawful-Competition contravening California law.
In this US-based case, it seems likely the fair use doctrine will be evoked; however we expect questions regarding AI and intellectual property rights globally to continue to dominate discussions across 2023.
What’s the Solution?
A path forward may be a techno-legal solution involving attribution and smart contracts that for example allow for royalty-type payments and consent notices, potentially triggered by the percentage of original work embedded in AI output.