Immediately hooked by the subtitle; literally my favorite data quote lol. I will say though, the entire enterprise of AI detection is fraught. The question that matters - as I see it, anyway - is whether the cost and burdens of False Positives are worth removing clear AI output like the 2nd half of this piece. Like, yeah if I see markdown headers I'd also be pretty confident it was AI generated, no modeling required 😂; so why harm folks who did put in effort by falsely labeling them as AI when the things that are clearly AI will be largely ignored? Is it for system automations? Well, like you demonstrated with the ByteByteGo piece, being AI-generated is not sufficient to qualify the utility of a piece. So what's really the purpose? Why do we want these detectors? Who does it serve?
Lol you are a true data soul :-D When everyone is whining, we’re experimenting.
It says 60% AI 40% human for me. Does everyone see the same?
I wonder if adding my edit with my screenshot changed the numbers? I'm now seeing 60/40 as well!
Immediately hooked by the subtitle; literally my favorite data quote lol. I will say though, the entire enterprise of AI detection is fraught. The question that matters - as I see it, anyway - is whether the cost and burdens of False Positives are worth removing clear AI output like the 2nd half of this piece. Like, yeah if I see markdown headers I'd also be pretty confident it was AI generated, no modeling required 😂; so why harm folks who did put in effort by falsely labeling them as AI when the things that are clearly AI will be largely ignored? Is it for system automations? Well, like you demonstrated with the ByteByteGo piece, being AI-generated is not sufficient to qualify the utility of a piece. So what's really the purpose? Why do we want these detectors? Who does it serve?
I was so hyped when that subtitle popped into my head haha.
Also, I'm super skeptical of the "how you wrote this" option. Seems like low hanging fruit to get human-labeled data for Pangram.