3.2M Records Show Proctored Math Scores Fell From 80% to 60% After ChatGPT
thisdudelikesAI · x · 2026-09-03
A UC Irvine PhD student, Sina Rismanchian, partnered with McGraw Hill researchers to analyze 3.2 million practice records from ALEKS, a math platform used by millions, spanning ten years around ChatGPT's arrival.
- Design: The platform naturally splits into word problems (pasteable into a chatbot) and graphing problems (must be done inside the software). Same students, same courses, two kinds of friction.
- Behavior shift: After ChatGPT, time on word problems collapsed — 31% less for high schoolers, 27% less for college students, while fifth graders barely moved.
- Diverging scores: Unsupervised practice scores went up, but proctored placement-test accuracy on similar items fell from 80% historically to 60% — a 25% drop in odds of a correct answer. Graphing performance didn't fall, ruling out general skill decline.
- Mechanism: Math skill is built in the minutes spent choosing a setup, guessing wrong, and fixing it. When the model hands you the path, you submit an answer but skip the part that turns exposure into memory.
- Implication: The author warns this likely extends to any task where the hard part can be pasted into a box — assume some learning is being copied away with it.
More from AGI Musings
- OpenClaw isn't dead: cheap Chinese models may push personal agents mainstream — FuSheng_0306 · 2026-09-03
- TMLR refocuses acceptance criteria to put more emphasis on clear writing — jordiponsdotme · 2026-09-03
- AI may matter most when it stops tuning experiments and starts inventing them — bravo_abad · 2026-09-03
- What do you call the backlash when you discover art was AI-made after the fact? — unsuspicious-account · 2026-09-03
- Guardian: Bill Simmons's ChatGPT podcast ad read is a breach of The Ringer's creative spirit — nordicinst · 2026-09-03
- Swedish blue-collar union chief economist: the fear is inefficient jobs dying too slowly, not too fast — Afinetheorem · 2026-09-03