Generative AI's negative impact on student academic performance exhibits a pronounced delayed effect — surfacing fully only after roughly two years of use — and the most capable students are the ones who suffer the greatest harm. This is not a problem inherent to AI itself but a problem of how it is used: students who "outsource" their homework to AI bear the heaviest learning losses.
I. Study Background
In June 2026, a joint team from Stockholm University and the University of Hong Kong published a study on SSRN tracking more than 26,000 middle and high school students in a county in central China over 30 months, from September 2022 to June 2025. It is the largest and longest empirical study to date on the effects of generative AI on K–12 education.
Eighty-one percent of students had used generative AI tools — with Doubao (豆包), DeepSeek, ChatGLM, Wenxin Yiyan (文心一言), and Tongyi Qianwen (通义千问) as the mainstream platforms. On the surface, AI users saw their homework scores improve by an average of 18% and completion time drop by 30% — but this is not learning; it is merely "looking like learning" by getting tasks done.
II. The Lagged-Damage Effect
The study's most central finding is this: the damage AI inflicts on learning ability is delayed in its appearance.
After approximately six months of AI use, monthly exam scores began to decline (an average drop of 20%). Two years on, in high-stakes progression exams, the impacts were severe: average scores fell by 24% in middle-school graduation exams (Zhongkao) and by 18% in the college entrance exam (Gaokao). Over the first two years, the overall average decline was only about 3.4%, meaning teachers would find it difficult to detect an anomaly against routine fluctuations — yet by 2025, when AI usage had reached 80%, the cumulative drop had approached 10%.
This "boiling frog" effect makes the problem nearly invisible within the window of time where intervention would be most effective.
III. The Decisive Role of Usage Mode
The study classified AI users into two categories:
- 81% "Outsourcing mode": Students directly had AI generate homework answers, chasing faster completion and higher scores. Homework was replaced by AI, with cognitive effort reduced to nearly zero.
- 19% "Augmentation mode": Students treated AI as a thinking partner, spending comparable time on homework to non-users, and their exam scores were largely unaffected.
The critical distinction is not "whether to use AI," but "whether cognitive effort is invested in the process of completing the task." Genuine learning depends on the mental work invested in problem-solving — precisely the part that AI most easily replaces, and the part it should least replace.
IV. Subject and Group Differences
Impact varied by subject: social sciences (politics, geography) declined 27% > STEM 22% > English 17% > native language (Chinese) 9%. The humanities, where AI can most directly generate complete answers, bore the greatest damage.
The group differences are especially telling: students who were already high-achieving suffered the largest drop (-24%), greater than lower-performing students (-16%). Researchers hypothesize that high-achieving students are more inclined to rely on AI for "optimal answers," forming a stronger dependency. Middle school students showed larger declines than high school students; boys declined 17% more than girls.
What this suggests is that AI is producing reverse selection — the group least meant to be harmed is being harmed the most.
V. The "False Fluency" Trap
The researchers identify a dangerous psychological mechanism: AI makes completing homework easy and fast, leading students to mistakenly believe their learning efficiency is improving in tandem, when in reality they are not genuinely mastering knowledge. This "false fluency" causes students to keep losing cognitive capacity under the illusion of high efficiency — until a major exam reveals the truth, by which time the damage is irreversible.
It is a systematic deception of learning motivation by technological convenience: rewarding behaviors that "look efficient," and penalizing the genuine cognitive effort that is truly demanding.
VI. Implications for AI in Education Policy
The policy implications of this study reach beyond education itself. When AI tool penetration exceeds 80% and the damage is delayed by two years before it becomes visible, education systems face a classic information-asymmetry dilemma: administrators need long-term data to judge AI's effects, but individual students experience daily, immediate positive feedback — "AI makes homework faster and scores higher."
"We need to pay attention to cultivating foundational cognitive abilities — information acquisition, sustained attention, task switching — and protect the brainpower investment that is indispensable in the learning process," rather than outsourcing thinking in the pursuit of efficiency.
VII. Limitations
- The study population consists of secondary-school students in a central Chinese county; generalizing the findings to other regions, age groups, and cultural contexts requires caution
- AI use in the study was self-directed by students, not experimentally controlled; causal inference must account for potential confounding variables
- The study period (2022–2025) coincided with the rapid proliferation of AI tools in China, and a "novelty effect" may have influenced the results
- As AI educational tools evolve in design — shifting from "giving answers" to "guiding thinking" — the observed effects may diminish
Core Insight
This 30-month longitudinal study reveals a counterintuitive reality: AI has not made learning easier — it has made the illusion of learning more convincing than ever before. When 81% of students use AI to complete homework, when their superficially improved grades mask a real decline in cognitive ability, the fundamental challenge facing education systems is not "whether to ban AI," but "how to redefine learning itself for a generation that has AI at its fingertips."