Thomas Basbøll asked his readers to grade an essay written by GPT-3. In Switzerland, I'd give it a 4.5 (4 is a pass, 6 is best). It's a list, not an argument: shuffling the order of points would not change much. It often uses placeholder language ("play a role") that leaves connections vague. While students handing in such essays might pass my course with a mediocre grade, AI needs to learn argumentation and avoidance of formulaic language to get them better grades. The former might be possible; the latter is tricky: mediocre human writers overuse formulas, and the AI presumably learns more from them than from good writers who avoid formulas. (Andrew Shields, #111words, 12 November 2022)
No comments:
Post a Comment