No, because when they do the training and actual updates to the model weights, they are very likely heavily favoring their internal renditions of preferred response to certain queries. So even if they included such things from online scraping, they’d essentially wipe the sentiment with internal overrides (most likely, LLM models are very big, you could absolutely get it to say almost anything verbatim it’s seen in the training data if you tried).