Scientists studied 26 thousand Chinese schoolchildren and revealed negative impact of AI

Economists from Stockholm University and the University of Hong Kong published a paper tracking how generative AI affects academic performance among 26,811 Chinese students.

The sample covers students from grades 7 through 12, meaning teenagers between twelve and seventeen years old, with panel data collected over thirty consecutive months.

The researchers pulled data from four sources, combining monthly closed-book test results, entrance exam scores for high school and college, plus homework grades and completion times across nine subjects.

Since students started using AI at different points, the team calculated the effect using a difference-in-differences method. This approach compares the same students before and after they adopted AI tools against peers who hadn't started using them yet.

The short-term picture looks great. After students began working with AI, homework grades rose 18% above baseline, while completion time dropped by roughly a third, from 64 minutes down to 45.

Then things start to diverge. Within six months, scores on monthly closed-book tests – the ones without any access to hints – fell by 20%. Entrance exam scores dropped 18% and 24%, with the full extent of the losses only becoming visible around two years after students started using AI.

The damage wasn't spread evenly:

• Social sciences took the biggest hit

• Math and natural sciences came in second

• Languages saw the smallest losses

• Younger students and boys stood out as particularly affected

• Students who were originally strong performers showed the steepest drops

Researchers tied the mechanism not to AI use itself, but to a specific pattern of behavior. Around 80% of students using AI tools appeared to simply hand their homework off to the AI, judging by the combination of unusually short completion times paired with high grades.

The rest of the sample told a different story. Students who spent roughly the same amount of time on homework as their AI-free peers saw only minor knowledge losses.