This is total speculation, but my guess is that human reviewers of AI-written text (whether code or natural language) are more likely to think that the text with emoji check marks, or dart-targets, or whatever, are correct. (My understanding is that many of these models are fine-tuned using humans who manually review their outputs.) In other words, LLMs were inadvertently trained to seem correct, and a little message that says "Boom! Task complete! How else may I help?" subconsciously leads you to think it's correct.

My guess is they were trained on other text from other contexts (e.g. ones where people actually use emojis naturally) and it transferred into the PR context, somehow.

Or someone made a call that emoji-infested text is "friendlier" and tuned the model to be "friendlier."

I suspect that this happens to be desired by the segment most enamored with LLMs today, and the two are co-evolving. I’ve seen discussions about how LM arena benchmarks might be nudging models in this direction.

AI sounds weird because most of the human reviewers are ESL.