Makes sense. The further away the target question is from the GSS training distribution, the more it relies on the ChatGPT prior. I assume if you click 'Investi

Makes sense. The further away the target question is from the GSS training distribution, the more it relies on the ChatGPT prior. I assume if you click 'Investigate Results', the confidence is 'low' and the most similar questions in the dataset are pretty far off