[peterth3]

How well GPTZero can detect ChatGPT generated text by measuring perplexity (how random the words choices are) and burstiness (how diverse the sentence structure is) shows that whatever algorithm our brain uses has stronger creative capabilities than this LLM.

GPT-3.5 isn’t a great writer like AlphaGo is a great go player. Maybe one day AI will generate better scripts and novels than humans, but not this model.

Medium-quality writing is ok for informative content though, but it’s problematic when the model doesn’t know fact from fiction. That’s the important complaint.

Is it dangerous? Maybe.

But is it useful? Not if it’s wrong too often.

You’re right that this tech should be taken seriously, but so should the hallucination problems. These problems can be solved. And maybe they should be solved before anyone trusts it with serious questions.

[tasuki]

> it’s problematic when the model doesn’t know fact from fiction.

This is in no way unique to an AI. Have you ever interacted with humans? Half the population thinks the other half can't tell fact from fiction. The other half thinks the same about the first half. We're all wrong about all the time.

[danaris]

There's a fundamental difference between "some people are wrong some of the time"—or even "half the population has trouble telling fact from fiction some of the time", if we grant that as true for the sake of argument—and "ChatGPT (and similar ML algorithms) don't even have a metric to determine truth from fiction; they just predict what's the most likely set of words to stick together in response to your prompt."

ChatGPT fundamentally cannot ever know when it's wrong. I should hope it goes without saying that that's not true of humans.

[tambourine_man]

From what I’ve read, LLM can’t reliably detect itself.