MIT apologizes, permanently pulls offline huge dataset that taught AI systems to use racist, misogynistic slurs • The Register

MIT apologizes, permanently pulls offline huge dataset that taught AI systems to use racist, misogynistic slurs • The Register

Image library with offensive tags taken offline

The Massachusetts Institute of Technology in Cambridge has removed from its servers a large database of images that were tagged with offensive terms, and has asked users to delete copies that might have been downloaded. Computer scientists Vinay Prabhu and Abeba Birhane discovered that the 80 Million Tiny Images database, created in 2008, contained images in which women and Black and Asian people were tagged with derogatory and vulgar language. Tagged images are used to train machine-learning systems, which then learn the biases that were in the tags, the researchers say.

The Register | 9 min readSource: arXiv preprint