Hacker News Reader

Comment by narmiouh | Hacker News Reader

What did you use to determine this?

Ask because many of the online tools I've tried, they will sometimes tag what I've written at 30-40% AI written and sometimes purely AI written stuff is flagged as 60-70% AI

View on HN

"Quietly". Also, the first paragraph. If this wasn't actually 100% LLM generated, one has almost certainly been used to imbue it with the flavour.

(That LLM rhetorical tic is especially funny because it's so obviously nonsensical in almost all contexts like OP - what would it mean for creatine supplementation's effects to not be 'quiet' and to be loud? The pills come with little Bluetooth speakers? You are trapped on the toilet for several hours a day with GI issues? The world suddenly takes on psychedelic colors and imagery all day long?)

As of now there are only two somewhat reliable tools.

First one is Pangram. Other available detectors are varying level of bad, with some of them entirely shit (eg zerogpt something).

The second one is human mind. Read enough of that slop and your brain hopefully will start detecting AI patterns.

And this article is totally AI, both to Pangram, and to my mind.

My experience in testing actual AI written content on willing participants is that people are entirely useless at detecting AI written content with any reliability whatsoever.

I don't know what your experience is, according to mine there are some people who are better than chance at picking this up.

And I believe my experience is something expected. People are also certain kind of a neural network. If an LLM system is trainable to be a decent detector, I don't see a reason why at least some people couldn't be.

I haven't seen any evidence an LLM is trainable to be a decent detector for anything people have made any kind of attempts at trying to get past them. Which is as expected as access to a detector effectively makes the problem equivalent to the halting problem (you can tweak the output using a detector as judge until you have a process to bypass it). Some of them are somewhat able to recognised "raw" output.

Yes, and the problem we are having here is 'raw' output. LLMgenerated slop is zero-effort bullshit, not an elaborate scheme to prove a philosophical thesis. There is no economy for mediaworkers doing the latter.

Similar as with coding, yes, halting problem!, but we've been always reviewing code nonetheless.

When I'm talking about raw vs. something processed here, the only processing I'm talking about is a prompt or two to clean up the obvious artefacts.

I'd be interested to test your claim, could you show me examples of this "prompt or two"?

> Read enough of that slop and your brain hopefully will start detecting AI patterns.

Like those people last week who ripped on an actual Monet thinking it was AI generated? No, the human brain only thinks it can pick up what's AI.

Read enough posts on LinkedIn and you’ll be a trained expert in spotting that slop.

Reply on news.ycombinator.com