Different models have different quirks. I use 4o at work and it has an annoying habit of taking things absurdly literally and using emojis and formatting too liberally instead of giving me a useful answer. Claude is much more subdued and helpful because it's conversational when you ask something open-ended, succinct when you ask it something data-driven, and almost always asks a follow-up question.

The sci-fi trope of being able to tell "AIs" apart from one another is absolutely coming true in real-time.

Might be stupid, but it works, for now. Over indexed tokens is another (common inclusions). Some grammar constructions, too, where over descriptiveness is present - though that's easy to read and probably a bit harder to code for.

If it works, it works.

It's a crude example, but pattern analysis to figure out who wrote a thing is an old, old technique; people have been doing it with Shakespeare stuff for centuries, in particular.

Why is it stupid? As long as people are too lazy to find/replace them, it seems to work.