Björn Ommer Featured in WirtschaftsWoche

On September 3, 2026, Germany’s Federal Court of Justice (BGH) heard for the first time a case on whether copyrighted photos may be used to create training data for artificial intelligence. A photographer had sued the non-profit organization LAION, which provides a large, publicly available dataset of image links and descriptions for research purposes.

In an article published by WirtschaftsWoche, MCML PI Björn Ommer, whose team developed the text-to-image generator Stable Diffusion in 2022, explains how such systems are trained. AI image generators learn from very large numbers of images paired with text descriptions, which allows them to connect words with visual features. In diffusion models, an image is gradually covered with noise, and the model learns to reverse this process and reconstruct the image. Ommer also describes how the LAION dataset was built from the Common Crawl web archive. To filter the image links and check whether each text matched its image, the images, or at least previews of them, had to be downloaded, analyzed, and then deleted. This step is at the center of the legal dispute.