Image-generation AIs memorise training images
High-profile AI image generators such as DALL-E and Stable Diffusion memorise images from the data they are trained on, ra ising concerns about potential copyright and privacy violations. Researchers at Google Deepmind, Princeton and other US universities extracted over one thousand training images from DALL-E , Google's Imagen, and Stable Diffusion , including photographs, film stills, copyrighted press photos, and trademarked company logos, and discovered that many of them were re-generated nearly exactly. The researchers got the models to 'nearly identically' reproduce over a hundred training images, often with hardly visible changes like more noise in the image, raising concerns about the reproduction and distribution of copyrighted material, as well as privacy risks to people who do not want their images being used to train AI. System 🤖 DALL-E Stable Diffusion Operator: Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramèr, Borja Balle, Daphne Ippolito, Eric Wallace Developer: Alphabet/Google; OpenAI; Stability AI Country: Global Sector: Multiple Purpose: Generate images Technology: Text-to-image; Generative adversarial network (GAN); Neural network; Deep learning; Machine learning Issue: Copyright; Privacy
- Date it happened
- 2023-02-01
- Organisation involved
- Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramèr, Borja Balle, Daphne Ippolito, Eric Wallace
- Product, system or model
- DALL-E; Imagen; Stable Diffusion
This incident was imported from AIAAIC and is used under CC BY-SA 4.0. Our additions to it — the structured fields, the translation, the checks against other reports — are published under the same licence.
This is a record of what was reported, not a finding that anyone broke the law. If it names your organisation and you believe it is wrong, the corrections process is free and open to everyone.