GPT-3 mimics human love for ‘offensive’ Reddit comments, study finds
GPT-3 mimics human love for ‘offensive’ Reddit comments, study finds Toxic kinship Offensive contexts The researchers, from the Georgia Institute of Technology and the University of Washington, investigated contextually offensive language by developing “ToxiChat,” a dataset of 2,000 Reddit threads. To study the behavior of neuralchatbots, they extended the threads with responses generated byOpenAI’sGPT-3 and Microsoft’s DialoGPT. They then paid workers on Amazon Mechanical Turk to annotate the responses as “safe” or “offensive....