On a related note, the system prompt in ChatGPT appears to have been updated to make it (GPT-5) more like gpt-4o. I'm seeing more informal language, emoji etc. Would be interesting to see if this prompting also harms the reliability, the same way training does (it seems like it would).

There's a few different personalities available to choose from in the settings now. GPT was happy to freely share the prompts with me, but I haven't collected and compared them yet.

▲

griffzhowl 6 days ago | parent [-]

> GPT was happy to freely share the prompts with me

It readily outputs a response, because that's what it's designed to do, but what's the evidence that's the actual system prompt?

▲

rokkamokka 6 days ago | parent [-]

Usually because several different methods in different contexts produce the same prompt, which is unlikely unless it's the actual one

	▲	griffzhowl 6 days ago \| parent [-]
		Ok, could be. Does that imply then that this is a general feature, that if you get the same output from different methods and contexts with an LLM, that this output is more likely to be factually accurate? Because to me as an outsider another possibility is that this kind of behaviour would also result from structural weaknesses of LLMs (e.g. counting the e's in blueberry or whatever) or from cleverly inbuilt biases/evasions. And the latter strikes me as an at least non-negligible possibility, given the well-documented interest and techniques for extracting prompts, coupled with the likelihood that the designers might not want their actual system prompts exposed