I remember reading a book a long time back, titled something like "non-linguistic analysis of call center conversations".
One main takeaway from the book was that "you can just look at the ratio of turn-taking duration, and which speaker/participant is spending how much time" to decide "what happened in the conversation".
The same goes for AI-generated conversations; verbose responses are the default behavior, and models are incentivised to keep that output-token ratio. Too easy to catch/notice, pretty annoying.
PS - I work in the Conversational AI space, and it is quite an effort to keep the ratio right for people to stay long enough on the phone with AI agents.
One main takeaway from the book was that "you can just look at the ratio of turn-taking duration, and which speaker/participant is spending how much time" to decide "what happened in the conversation".
The same goes for AI-generated conversations; verbose responses are the default behavior, and models are incentivised to keep that output-token ratio. Too easy to catch/notice, pretty annoying.
PS - I work in the Conversational AI space, and it is quite an effort to keep the ratio right for people to stay long enough on the phone with AI agents.