Stanford Researchers Find 3,226 Instances of Known Child Sexual Abuse Material in LAION-5B — The Dataset Training Stable Diffusion, Midjourney, and Imagen. LAION Takes the Dataset Offline. The Models Already Trained on It Still Exist and Still Run Locally.
Dec 20, 2023: Stanford Internet Observatory cross-references LAION-5B image hashes against NCMEC's CSAM database. 3,226 confirmed matches. The nonprofit LAION pulls the dataset the same day. But Stable Diffusion 1.5, Midjourney v4/v5, Google's Imagen, and every fine-tune ever downloaded already has the representations baked into their weights. There is no way to 'unlearn' specific training samples from a diffusion model. Researchers: 'The horse has left the barn.'
Stanford Researchers Find 3,226 Instances of Known Child Sexual Abuse Material in LAION-5B — The Dataset Training Stable Diffusion, Midjourney, and Imagen. LAION Takes the Dataset Offline. The Models Already Trained on It Still Exist and Still Run Locally.
Dec 20, 2023: Stanford Internet Observatory cross-references LAION-5B image hashes against NCMEC's CSAM database. 3,226 confirmed matches. The nonprofit LAION pulls the dataset the same day. But Stable Diffusion 1.5, Midjourney v4/v5, Google's Imagen, and every fine-tune ever downloaded already has the representations baked into their weights. There is no way to 'unlearn' specific training samples from a diffusion model. Researchers: 'The horse has left the barn.'
Weirdness Classification
10/10 — Deeply unhinged
Know something weirder?
Submit your own AI incident report to the public record.