zip

chaos gf

@zip

  • she/her

annoying @vogon since 2014. still gay tho.
low energy software menace. i like taking photos, learning weird programming languages and being on the internet.

log inask
profile picture by
cohost.org/mleelunsford
zip
@zip
possumskull
@possumskull
bark
@bark

detecting ai-generated papers through “tortured phrases”

there's a paper on the arxiv [1] that discusses detecting ai-generated research papers through so called “tortured phrases”; that is, phrases which resemble standard field-specific jargon, but seemingly mangled by a thesaurus.

they built a set of these tortured phrases (the table above) and searched the literature. and found. so many examples.

anyway.

tag yourself i'm “underground creepy crawly province”

[1]: G. Cabanac, C. Labbé and A. Magazinov, “Tortured phrases: A dubious writing style emerging in science,” Jul. 2021, arXiv: 2107.06751v1. Available: https://arxiv.org/abs/2107.06751v1

belarius
@belarius

BORN TO credulous Bayes

FORCED TO mean supreme blunder

zip
@zip

I'm so excited for the tiktok kids to use this sort of thing to bowlderize their way past filters. just generating absolute nonsense like "he committed effluent drain(sewer) helter skelter(slide)" or "they should check his robust(hard) excursion(drive), there's no way he's not a gender(sex) evil-doer(offender)"

You must log in to comment.

in reply to @bark's post:

gull
@gull

they used to do this shit with shady APK download sites, and probably other shady download sites too, presumably to try and dodge guys sniffing out said shady sites. this led to deeply deeply funny manglings of descriptions

anyway. haze figuring

login to reply
upthorn
@upthorn

A lot of these seem like reasonable mistakes a foreign language speaker could make. Or machine translation of a human-written foreign language article.

As soon as I read the first line I was reminded of the time a Chinese professor in an intro CS course taught us about the "deadly embrace" (deadlock) that occurs when an I/O process is waiting on something that is waiting on that I/O process.

login to reply
RenaKunisaki
@RenaKunisaki

These seem like less of a way to detect AI specifically, and more of a way to detect plagarism.

It's a basic trick: you copy existing text, then find-and-replace with synonyms to avoid automatic detection.

This is often done poorly, resulting in book reports about how Enormous Sibling is Viewing You, or strategy guides telling you to take on the Pokemon Recreation Center before you fight Team Missile.

login to reply

in reply to @zip's post: