As an educator, I've come across a chart claiming to show common ways students use ChatGPT. I'm curious about how OpenAI tracks prompt usage. I know they monitor for alarming trends, but do they really save and analyze the prompts to the extent that someone can query and find details about user prompts? This chart suggests that they might be promoting a narrative that students aren't using it for cheating, as it doesn't show any prompts that directly ask for solutions. Is this typical behavior for OpenAI, and is there reason for me to be skeptical about the validity of this information?
4 Answers
It sounds like OpenAI uses user data to fine-tune and improve their models, which is why there's the option to let your prompts be included. They create analytics from aggregate data, but the specific queries on prompts aren’t often discussed publicly. The focus is on enhancing overall performance while maintaining user privacy.
Sure, the chart claims it's from university students, but let’s not forget that 'favorite prompts' can often be biased feedback. Polls like this can be very self-selective, which might not represent the whole picture of how students are actually using the tool.
Right? If those cheating students were voluntarily replying, that would be surprising!
The whole idea of improving models comes from user interactions, but it’s also linked to OpenAI's 'free' version. They do collect data, but it's meant to benefit everyone without compromising user privacy. Still, how they present this data can raise questions.
I see your point. They should clarify how they analyze the data, apart from just showing a pretty chart.
Regarding that chart, it may have originated from a self-reporting poll among students, which can skew results. Students who are using ChatGPT to cheat likely wouldn't admit it in a survey, so the data might not be as reliable as it looks.
Exactly! It's hard to believe that the ones who are cheating would volunteer that info in a poll.
I always thought that meant they used all chats collectively for training, not specific data queries. They say it's to improve the models, but I'm wondering if they publicly address how they analyze data for insights.