Hacker News Reader

Comment by kredd | Hacker News Reader

Counterpoint - confirmation bias. If you make the LLMs respond with specific tones, mix in some casual grammar mistakes, I don’t think you can distinguish it.

View on HN

Sure, but that feels like a lot more work than most bot makers are putting in. I'm not claiming folks have 100% hit rate on detecting bots, but the run of "disregard instructions" memes have shown a good amount of pretty low effort bots. It takes a lot of work, patience, and practice to make a bot that feels human. I don't doubt that there's influence operations out there that are putting in the effort, but just like I still get spam email about V1AGRA multiple times a day, it's not clear the cost benefit is worth it for scammers and influence operations to pay the cost in LLM tokens to do that just yet.

Do you have any input on anything after my first paragraph?

Can you prove you are not a bot? Your pseudo is SUS.

People think they can recognize an LLM, but what they're actually recognizing is the ChatGPT default prompt.

Just a thought, are there any spell-checkers-backwards? They take some good prose and introduce some common mistakes?

they're all tainted with GPTisms, which are trivial to spot.

I’m a bit surprised with the reaction in the comments section. My genuine belief, and beliefs of what I assume the bubbles I’m part of, LLMs are not detectable by any means. And I assumed the problem is, the less technical people have not been made aware enough in real life, so they’re still not skeptical.

It was a bit hard for me to come into terms with, in the last 7-8 years, too. I remember my 2000s, where talking to anyone online or playing with them would guarantee that there’s a real person behind it. Or at least the chances were incredibly high. Now? I assume the chances are incredibly low. Or at least extreme financial incentives for online interactions. Which ruined the internet somewhat for me. Which, I think, will get worse. Eh, good old days.

Reply on news.ycombinator.com