[malauxyeux]

> We can all prompt ChatGPT.

I'm largely blind to this. If they're not wordy and overly polite, pasted ChatGPT answers just don't appear on my radar.

Can someone help me out? What points to this answer in particular being from ChatGPT?

[scrollaway]

Some tell-tale signs (I work with GPT a lot as I mentioned in a sibling comment):

- It's a list, it's numbered, it's got pretty consistent markdown formatting. This is especially present with ChatGPT

- Title Casing In Each List Item

- Strong usage of the passive voice

- Strong CYA tone. "might promote", "studies suggest", "has been associated", "can convert", "possibly lowering", "thought to increase"

The general structure is very consistent with GPT too. Once you've seen a lot of sessions it's just... plain obvious. Especially if you step back and think: "Would people actually.. write like this?"

[mdp2021]

Sure, but hints are not proof. In fact,

> Would people actually[] write like this?

: some people do. What you call «CYA tone» can overlap with what for others is "precision".

It is the quality of the text, then, that hints further to actual intelligence or "artificial struggle".

[scrollaway]

GPT4 can generate some extraordinarily high quality text if you know how to prompt it. But this ain't it - it's some of the most boring way to prompt. The OP's response is what happens when you prompt with some article summary and a "What are some possible ways we could increase the production of nitric oxide in the body".

And no, that trashy CYA tone is not "precision", if anything, it's vagueness. It's weasel words.

[mdp2021]

> no, that trashy CYA tone is not "precision", if anything, it's vagueness. It's weasel words

That is not what I said. /That/ «trashy CYA tone is not "precision"». But some use similar expressions to those you noted in order to be factual.

Some texts give a strong impression of fakery; some texts could give a wrong impression upon brutal use of raw Bayesian indicators. Signs orient, do not decide. Hints are not proof.

Some patterns in LLM output can be caricatures of proper efforts (factuality for precision, when relevant, is one of them).

So, people may write similarly to that. (Only, hopefully, well beyond veneer.)

[addisonl]

Well, considering it was trained from content created from how people actually write, yes?

[scrollaway]

Ah yes, just like how I know most of my american friends have exactly 0.78 kids.

[mpalmer]

Brief summarization that seems redundant to the topic because it's the model priming itself to generate a response in the requested form, which is usually a:

3-10 item list which mentions certain phrases related to the prompt every time ("nitric oxide" here).

Disclaimer about accuracy/expertise.

[JPLeRouzic]

So actually, it's a well formed answer?

[mpalmer]

It very likely bears a close resemblance to a member of the theoretical set of all well-formed answers.

Problem is all the other sets it belongs to, including the set of all answers containing unverified factual statements, not to mention the set of answers that resemble the output of a large language model.

[nannal]

Chatgpt loves lists like this, then length and "tone" can be the most obvious indicators. You can modify those attributes through prompting but that should help find unmodified output