[mintaka5]

but is quality of an essay a artificially derived standard? i mean we can talk about how great the material is that is produced, but seriously, folks, aren't we chasing our tails here? is an AI good at wirting essays because of its own standard of quality or is it being measured against our own already established low bar? i mean if we're saying the computer is better than us, measured against a standard we dreamed up and supplied, it really is lacking objectivity here.

[jfengel]

Precisely. This paper used human graders, but these kinds of essays are also being graded by AI. The essays use a standard form, and they're marked objectively against it.

AI is very good at writing things that don't need to be written. It regurgitates, just as students are expected to do. The essays aren't expected to be good, for the same reason nobody makes any math discoveries in their homework. They're expected to go on and eventually do something of actual merit.

That's harder to grade, but thus far there's no indication that AI will develop the way humans do. Lousy student essays hopefully give way to good writing. AI is trained to do essays, but nobody wants essays.

Maybe this is a waypoint for AI as well. I've got my doubts that the tactics we're using can ever develop that way, but I'll try to keep my mind open in case something materializes -- a pinch of randomness that better mimics creativity and insight.