What actually makes an interview useful — turns out it's not what researchers thought

Researchers built a dataset of 343 real interviews with 16,940 responses and tested 10 different measures of interview quality to see which ones actually predict whether a response helps answer the research question. It turns out direct relevance to the research question is the strongest predictor — but two measures that interview systems commonly use (clarity and informativeness) don't predict usefulness at all.

For decades, qualitative researchers have used proxy measures to judge whether an interview response is good — clarity, detail, surprise value. This paper shows those proxies are decoupled from whether the response actually contributes to the study's findings. That means automated interview systems built to optimize for clarity or informativeness are optimizing for the wrong thing. The real signal is relevance to the specific research question, which is messier and harder to automate but actually matters.