The tendency of LLM outputs to vary significantly based on small changes in how a request is phrased.