GPT models are predicting the next token with a certain probability. If Google has access to the data-set used for training a specific model and if it has access to the AI model itself, it can calculate that probability and identify the content that has been generated by that model.
BUT if the AI model has been trained/ fine-tuned on a private data-set, there is no way for Google to differentiate the generated content from the human written content.
My personal opinion is that Google will use their AI to improve other aspects that will help them establish if the content is good (eg: did the content deliver what the reader was expecting? could this content be better?)