Why AI adoption severs the link between task performance and learning, and what that means for evaluation, product, and governance
What Homework Can No Longer Tell Us
I recently had the opportunity to review the results of a new study from Stockholm University and the University of Hong Kong on the use of AI in education. The study followed 26,811 Chinese high school students for 30 months. The data at the end of the study is quite revealing: the improvement in homework completion thanks to the use of AI was associated with a decrease in test scores where they could not use AI.
What the data says: after the adoption of generative AI, students’ homework grades increased by 18%, while the time spent completing them decreased by almost a third. However, although these numbers might seem expected for an educational product, when the exams arrived where these students could not use AI, their grades dropped by 20% in six months. Moreover, two years later, when students faced China’s most important university entrance exams (Zhongkao and Gaokao), scores dropped between 18% and 24%.
In short: the assignments improved, the students didn’t.
The Metric Never Measured What We Thought
At this point, we might ask ourselves whether the use of AI helps or hinders student learning. However, upon closer examination of the data, we see that the majority (81%) of students who completed assignments in less time thanks to AI were the ones who subsequently received the lowest grades. Conversely, the remaining students who used AI but dedicated a similar amount of time to their assignments as those who didn’t use it, received similar grades.
In other words, the key variable is not so much whether AI was used or not, but rather the time spent completing the assignments.
And here lies the fundamental problem. For years, homework grades have served as a leading indicator: a frequent signal that allowed us to anticipate the results of the much less frequent exams. But the use of generative AI has broken that relationship because, as the study shows, the grades on homework completed by most students are no longer indicative of how they will perform on exams.
What does this mean in practice?
For assessment design, this doesn’t mean we should prohibit the use of AI in homework. It means that homework grades, if AI is being used, should have less weight and be complemented by a second variable that comes into play: the time spent on the task, measured rigorously, and not as an accidental byproduct of a submission timestamp.
For educational product roadmaps, this forces us to rethink what “engagement” means in an AI-powered tutoring role. A tool that is only optimized to reduce homework time and raise grades is, according to this data, reproducing the very pattern of failure. Therefore, perhaps the design question we should be asking ourselves is not how quickly a student arrives at the correct answer, but whether the product can distinguish between a student who worked to arrive at that answer and one who didn’t work at all, and treat those two cases differently.
For governance, this is a type of risk that current frameworks are not designed to detect. Checklists on bias, transparency, and accountability ask whether the outputs of a system are fair and explainable, not whether the underlying process is still functioning as expected. Perhaps a new risk category is needed for this.
Conclusions
If we look closely, AI did not break students’ capacity to learn. What it has done is make it more difficult, based on the variables we usually use to assess them, to know whether they have learned.
This connects to a detail revealed by the study: teachers didn’t notice what was happening because each one only teaches one subject, and in each subject individually, this could go unnoticed. Parents did notice because they saw the entire range of subjects, but they had no way of linking the drop in grades to the use of AI months earlier.
Some information may be outdated
