pagenet

I’ve only been trying for a couple days, but I found you can fool 1/2 the detectors with creative prompting alone .. when you add a spinner to it, you can fool most all of them .. WITHOUT leaving a footprint that makes it obvious you are trying to bypass detection.

SpinRewriter’s humanize function is a bit weak .. but it does help.   You can add grammar mistakes, but I found that doesn’t help the detections scores as much as you think it should.

I found Claude 2 to be the best at passing detection, especially when you add some creative prompting and a spinner .. but believe it or not, creative “write as a human” prompts alone appear to do most of the bypass.

WORKFLOW
Creative “Humanized” Prompt -> Claude 2 -> Spinrewriter Synonyms -> Spinrewriter Humanize function -> Then Fix Grammar

So you can have a perfect grammar and still pass as human on just about every free detector (@ http://www.undetectable.ai).   Not sure about Originality or Wilson.

I found it also works pretty well on OpenRouter LLMs .. even the free ones … This formula works right now, but who knows in the future.

This WHITEPAPER seems to have a way to solve it for good though.
https://arxiv.org/abs/2305.10847v5
https://arxiv.org/html/2305.10847v5

Python Github is here
https://github.com/ColinLu50/Evade-GPT-Detector

I don’t know much about Python, but if you do, you might want to check it out and see if you can convert it to PHP for your plugin, or better yet .. sell it as an addon rather than another API.   API costs are prohibitive to scaling   ;

Keep at it. .. we’re getting close.

Thank you for your quick answer.

So, I noticed on line 1548 the following code:
if(str_word_count($html) >= 2500)

What happens if word count is over 2500?   Does the rest of the post get truncated or just not spun?

PS: I noticed that SpinRewriter sample SDK code does not have 30k limit anymore.   It’s now based on time (7 seconds between API requests / 500 per day max).