Bramblethorn
Sleep-deprived
- Joined
- Feb 16, 2012
- Posts
- 17,950
Problem is that any simple rule for detecting AIs can be used to train the AIs not to do that thing. So we're left with the complicated ones.This. We need to get clarity on the trigger phrases and/or sentence structures, so we can compile the differences between human writing and derivative word prediction software (let's not call it AI, because it isn't).
I think we can all recognise AI when we read it, but reducing it to the "reasons why" does get a little harder.
This comment hit me right in the gut. I had my writing flagged by AI and I know I wrote it. I just put it through several AI detectors and was amazed to see what was flagged. Even sex scenes were flagged and chatgpt won't even write those!
It's not supposed to, but it can be tricked by things like the "napalm grandma" exploit. And there are other, similar products which aren't as restricted as GPT.
I still don't put stock in AI being able to write. I tried several of those systems. They have the attention span of a child, or worse than one. They can't keep the line of thought for more than a paragraph.
They have gotten significantly better in the last year or so, though there's still a huge amount of hype and wishful thinking. I'm not aware of anything that can write a story that holds together for more than a few hundred words, but they seem to be used more often as a filler tool - a human will write an intro or an outline, and get it to fill that in a paragraph at a time, with the human editing as needed.