I read this post (and then the paper) scratching my head over two details: which OpenAI model was this using (it just said "ChatGPT") and was it using ChatGPT's data analysis mode, aka ChatGPT Code Interpreter.<p>I finally found this tucked away towards the end of the paper:<p>"Browser-based version of ChatGPT version was updated as we conducted our task-based studies. P1-P9 used the March 14, 2023 version of ChatGPT, and P10-P14 used the March 23, 2023 version."<p>GPT-4's release date was March 14th 2023 - but only for paid users. Was this study using GPT-4 or GPT-3.5? That's still not clear to me.<p>Even more importantly: was it using Code Interpreter / data analysis mode? <a href="https://help.openai.com/en/articles/8437071-data-analysis-with-chatgpt" rel="nofollow">https://help.openai.com/en/articles/8437071-data-analysis-wi...</a><p>I'm almost certain it was not - none of the screenshots showed evidence of that mode, and that feature was made available (as part of the plugins beta) to beta testers on March 23rd - it's weird that the dates mentioned in the paper happen to coincide exactly with the GPT-4 and the Code Interpreter preview dates!<p>Without including data analysis mode I don't think this paper addresses the question "can ChatGPT do data science?". That's what that 15 month old feature is for, and any study that doesn't include it is missing the most important factor in answering that question.