What's this about?
A new study on student AI use in China produced a chart that circulated widely: the long-standing relationship where good homework grades predicted good test grades has weakened, and may have inverted. Ethan Mollick supplied the crucial context — this is real, but the pre-AI baseline was already bad. Doing homework improved final test grades for 86% of college students in 2008, but only 45% by 2017, because students were copying from the internet.
What exactly does an inverted correlation mean here?
Homework was historically useful as both practice and signal: doing it made you better, and doing it well predicted you'd do well on the test. If that correlation goes to zero, homework has stopped measuring anything. If it goes *negative* — good homework grades predicting worse test scores — something stronger is happening: the students producing the best homework are the ones outsourcing it most completely, and they're the ones learning least. The assignment has become an anti-signal, which is worse than a useless one, because institutions still act on it.
What's the community saying?
- Mollick's "it was already bad" framing is the most important correction, and the 86%-to-45% decline between 2008 and 2017 is the number that reframes the story from "AI broke education" to "AI accelerated an existing collapse."
- The inversion is what makes the chart shocking, and also what warrants caution — an inverted correlation in one study is a strong claim that deserves replication.
- It sharpens the assessment question rather than the cheating question. If unsupervised work no longer measures learning, the response is redesigning assessment, not detecting AI — which connects to this week's watermarking and AI-detection threads.