Token Prediction as Implicit Classification to Identify LLM-Generated Text