Hacker News new | past | comments | ask | show | jobs | submit
SOTGO | on: Failing grades soar with AI usage, dwindling math skills in Berkeley CS classes
I think curving has its place. One of my math professors explained that in his opinion an effective test should differentiate performance as much as possible. The top students should score very well and the bottom students should score very poorly. If all the scores are clustered near the top (>80% for example) then it's hard to tell who really mastered the material and who just muddled through. Then, once you've sorted the students you can apply an appropriate curve. He did not have pre-defined thresholds, for each exam he would evaluate when he felt like the quality of work changed from an A to an A-, A- to B+ etc. The curves were very fair; he wasn't trying to force some number of As Bs or Fs, but it did increase my stress levels not knowing in advance how well I needed to do on each exam
jvanderbot | parent
Yes, curving has a place here - and it is to evaluate, as you put it, whether the test differentiates performance as much as possible.

If you curve the students after the test, you are applying subjective edits to the graded performance just so the distribution of grades matches the measure of your tests effectiveness. That's just hacking the metric.

Further, even if you believe that tests should differentiate mastery (not students), your test should have teased out the differences or given you enough confidence to provide As to everyone who mastered the material - which should be absolutely possible! There's no a priori reason that all students cannot absolutely get the same grade, except for the a priori assumption that grades are for differentiation of students themselves (this year's A means this is the best student of this year), vs indicating mastery (all students absolutely crushed this exam).

You can dock points for style, or unnecessary struggle, or whatever subjective metric you want, but fudging the grades based on vibes to fit a prior-assumed distribution is just kinda "test effectiveness laundering"