@timnitGebru to make it even worse, the false-positive rate of these "AI detectors" is much higher for non-native English speakers:
@jamesmarshall @timnitGebru
That makes sense. The kind of writing that AI does, especially if you don't lead it into any particular style, tends to be that sort of dry read that non-native speakers end up writing, with similar mistakes, possibly because it's been trained on other non-native speakers making those same mistakes: I'm certainly impressed at how well these supposedly English AI can handle switching into Spanish or German for me. I also wonder if the fact that translation software is now AI driven means similar patterns will show up.
Once again, minorities get disproportionately affected by these Tech Bros.
@jamesmarshall @timnitGebru
Actually, you know what also makes sense? They are using cheap labor in the third world by hiring non-native speakers to generate example text to train these things off of, and those people are writing their prompts in a way common for non-native speakers. Another theory would be the point out that most text is probably written by people who aren't from the main Anglo-Sphere of the US, Canada, and Britain.
I'm just fascinated by the number of problems with this entire model.
Pretty sure though that part of the grift here is to make a text AI, claim that it can write like any human, then get you to buy their software specifically made to catch their AI. Whether or not the AI or the AI detection software actually works reliably is irrelevant, as long as they can get you to pay for it.
Timnit Gebru (she/her).
James M.
Raccoon🇺🇸🏳️🌈