I think they can run benchmarks to see how likely it is for prompts to return exact copies of their training data and use those benchmarks to help tune their training procedures.