We introduce an automated method for evaluating the quality of AI-generated text by repurposing prior human evaluation data. The method is called Near-Negative Distinction.