26,000 Student Study Shows AI Hidden Learning Costs Take Two Years to Surface
- Ethan Carter
- Jul 5
- 3 min read
A panel study conducted by researchers at Peking University’s Graduate School of Education that tracked 26,000 students in grades 7-12 for 30 months found a clear pattern. Homework grades rose 18 percent and completion time fell from 64 minutes to 45 minutes when AI tools were used. Closed-book exam scores dropped 20 percent overall and college entrance exam results fell between 18 and 24 percent.
The performance decline did not appear immediately. Researchers recorded the first measurable drop only after roughly two years of regular AI use. Early gains averaged 25 percent lower exam results in the first year; the gap narrowed slightly to 16 percent after that but never disappeared.
Social studies showed the largest decline at 27 percent, followed by STEM subjects at 22 percent. English scores fell 17 percent and Chinese language scores fell 9 percent. Students who used AI more than five hours per week lost about 30 percent in long-term measures, while one hour per week still produced a 5 percent loss.
Eighty-one percent of long-term AI users finished homework in under 50 minutes, a threshold the study treats as a strong signal of external assistance rather than independent work. The data came from a continuous panel that recorded tool usage logs from major platforms including ChatGPT and similar generative tools, assignment timestamps, and scores on standardized tests such as national college-entrance examinations every semester.
The study design followed the same students across three school years, allowing direct before-and-after comparison for each individual. Researchers controlled for prior achievement, family income, and school resources. The two-year lag remained significant after these adjustments.
School administrators have begun reviewing homework policies in light of the findings. Several districts now require at least one closed-book assessment per month that prohibits AI access. Teachers report increased difficulty distinguishing independent work from assisted work when time stamps show rapid completion.
Parents have received mixed messages. Some welcome the immediate efficiency gains on daily assignments. Others worry that the delayed test-score impact will affect college admissions years later. The study authors recommend limiting weekly AI time on core subjects rather than banning tools outright.
The same dataset showed that short-term homework improvements did not translate into better conceptual retention. Students who relied on AI for social studies essays scored lower on map and timeline questions that required recall without notes. Similar gaps appeared in algebra and chemistry problem sets.
No single AI tool dominated the usage logs. The pattern held across multiple platforms once weekly exposure exceeded three hours. The researchers noted that the effect size grew with cumulative exposure rather than with any specific product feature.
School leaders now face a timing problem. Policies that only measure homework completion will continue to register success even as exam results slip. The two-year delay means current seventh graders will reach high-stakes tests before the full cost becomes visible in aggregate data.
Intervention timing matters. The study found that students who reduced AI time after 18 months recovered roughly half of the lost exam performance by the end of year three. Recovery was weaker when reduction began later. The authors therefore advise earlier limits rather than waiting for test results to confirm the problem.
Future data releases will track the same cohort into grade 12. Observers will watch whether the 18-to-24 percent entrance-exam gap narrows, stays constant, or widens. Districts that introduced monthly closed-book checks will serve as a natural comparison group. Any measurable difference in those schools could indicate whether the two-year lag can be shortened by changes in classroom practice.
