How to generate AI articles that are 99.98% *undetectable* by AI output detectors.
Hint: You just tell it to write like a human ⬇️
Hint: You just tell it to write like a human ⬇️
Here I have told AI to write an article about cheese... 
HuggingFace GPT-2 Output detector believes with high certainty (99.98%) that it's fake. And rightly so...
But what about if we tell the AI to write in a 'human' way, with unlikely and nuanced sentences?
The output detector now believes this is 99.98% real.
For what it's worth, this needs a lot more testing, but with the right prompts I think it can be versatile enough to bypass detectors.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
