Text this: Impact of large language models on peer review opinions from a fine-grained perspective: Evidence from top conference proceedings in AI