OpenAI admits that AI writing detectors don’t work
OpenAI admits that AI writing detectors don’t work
No detectors "reliably distinguish between AI-generated and human-generated content."
You're viewing a single thread.
"admit" like this was something they denied? Everyone always knew that ai text detectors don't work and they never claimed otherwise.
The issue is that schools have been using detectors to flag AI essays. When students (wrongly) get caught up in them, they get penalized even if they never used any AI to help them write the essay in question. Sort of like a plagiarism filter falsely flagging a paragraph as plagiarized, even if the student didn’t plagiarize it.
It also disproportionately flags the work of neurodivergent students, to add a bonus reason that these detectors are a dogshit idea.
Not that I don’t believe you, because I do, but do you have a citation for this?
Neurodivergent people are secretly robots, and sometimes the AI text detectors can pick up on this, and risk ruining our cover.
In all seriousness, I looked it up out of curiosity and couldn't find a study stating that, but that isn't surprising as the use of AI detectors is relatively new. I do think there is a high likelihood that the statement is true, just due to how a neurodivergent person often writes compared to a neurotypical person. This is not something that you could say matter-of-fact though, just anecdotally.
I’m glad I went to school (mostly) before all the automatic scanning..
In college I had a lot of stuff flagged as plagiarism, including a metaphysics paper I wrote in which I created a new hypothesis for consciousness (not necessarily a good one, mind, but entirely unique) because yup, robot.
It’s so weird that writing properly and without error gets you flagged as a cheater…. That supposed to be the damned standard you are being measured on..
But yeah, it sounds highly plausible, but I don’t tend to take assertions as fact without support, so thanks for looking into it for me :)
From what I understand, a lot of AI text detectors work from measuring "perplexity and burstiness" which basically means, lower randomness, lower emotion, and sentence uniformity is more likely to get flagged as AI text. Those are all things that can be associated with neurodivergence, so I see where that statement would come from. That's also the way you are expected to write formal essays.
As for plagiarism detectors, I don't know much about them except that they false flag the shit out of everything.
Oh yeah it's an issue, but none of that is on OpenAI. There's no admission here. It's a statement from an authority to shut up the idiots, like a map maker saying earth is a sphere, something we already know but somehow it's still believed by many.