Empire of AI by Karen Hao, p. 136
«For years, the tech industry had relied on poorly paid workers in precarious economic conditions to perform essential data preparation tasks for its AI models, such as categorising text and labelling images. Soon after GPT-3 normalised the use of giant, poorer quality datasets, the demand for the work shifted from the handling of largely benign content to frequently disturbing content, including for the purposes of content moderation, much like social media before it. Such moderation was necessary to prevent generative AI systems from reproducing the most vile parts of their all-encompassing datasets — descriptions and depictions of violence, sexual abuse, or self-har — to hundreds of millions of users.»