zlacker

>>djtrip+(OP)
Can generated responses even be reliably detected? How does that work?

>>jfoste+ms1
For all but the shortest fragments of text, I suspect the answer to this will be "yes" and this will be achieved in a short timeframe. Interact with Chat-GPT enough and you soon begin to tire of its middle-of-the-road prose. Because it is in some sense a statistical average of its inputs, its writing feels very monotonous.

Of course, you can try to push it in the direction of something more interesting... "write in the style of X" so that it will regurgitate from a subsection of its training input, or push it onto a niche topic. Across a wide range of inputs, though, I predict it will be easily classified.

As training GPT will take far longer than training a GPT-detector, this is an arms race that the detectors will win for the most part. Anyone wishing to use GPT to generate middling essays to cheat at school will have to spend a lot of time to disguise the text (perhaps repeatedly running against open-source detectors) and then take the risk that detection does not improve between their submission and review.

>>random+bO1
If that's so, where is the GPT detector?

>>jfoste+pY1
mmm well how long did GPT take? Say a quarter of that time?

>>random+OB3
OpenAI started late 2015. I don't know when they started working on GPT, but GPT-2 was early 2019.

Nearly 4 years later, still no GPT-2 detector?