Yes, that was intentional due to IRB issues here. If you email us we can send you some examples!

So, a good lesson here. ChatGPT does not introspect -- it only knows how texts in its training corpus have answered questions about behaviors. Since questions about language are typically about humans, it will answer language-related questions as if it were human. 6/6

Then, I asked it why it didn't give me that answer initially. It replied that in its training corpus, this type of question is typically asked of a human. So, naturally, it responded assuming a human was the addressee of the question. 5/?

I was finally able to get an answer about why an LLM style-shifts when I asked it explicitly to think about why an LLM like ChatGPT would style-shift and not a human. 4/?

Then I asked it what source of information it was using to make these claims. It turns out, this response was based solely on its understanding of how humans style-shift. 3/?

The Singlish disappeared and was replaced by American slang ("lowkey", etc.). When I asked ChatGPT what had prompted it to shift, here's what it initially told me. 2/?

Students report that they self-police by avoiding terms and phrases that they believe are associated with ChatGPT. #LSA2025

The corpus includes an interactive map that shows the locations of speakers and how country borders shifted over the course of the corpus, which is helpful in ascertaining what the language of school would have been for corpus speakers and other info. #LSA2025

The authors are working on developing a transcribed corpus of spoken Yiddish from a subset of the corpus held by the USC Shoah Foundation, which currently has a high barrier of access and does not make files available for researchers. #LSA2025

Cisgender listening subjects who are unfamiliar with trans people expect trans voices to differ from cis voices, based on top-down stereotypes. Stereotypes for trans people and speech differ between cis and trans communities, resulting in differences in perception. #LSA2025

Really is preferred for "good-type" adjectives (really good) rather than "tall-type" adjectives. Also a main effect of age, with younger speakers preferring really. #LSA2025

Greatest awareness of aw Canadian raising, then ay monophthongization, with bag raising and ay raising at much lower levels. #LSA2025