Surely because LLM generated code is part of the training data for the model, so code/patterns it can work with is closer to its training data.