Evidence and limitation
A large study found a gap between homework productivity and unaided performance
CEPR Discussion Paper DP21577 tracked 26,811 Chinese secondary-school students for 30 months. It reports higher homework scores and faster completion after AI adoption, alongside lower closed-book and entrance-exam scores for the observed population. The paper is evidence about one setting and pattern of use, not proof that every AI-assisted learner will have the same outcome.
How we evaluated this
Keep the source, method, and boundary visible.
The numbers on this page come from the paper's public abstract, not VUST user data. Read the CEPR paper record for authors, methodology, population, and the reported estimates.
- Best for
- Revision, exam preparation, and any task where you must later explain the method unaided.
- Not ideal for
- Copying a finished assignment or relying on a generated answer without checking it.
- In short
- Use AI to create the next learning action, not to remove the action altogether.
See the difference
The failure mode is not using AI; it is outsourcing the attempt that you need to be able to repeat later.
A four-step study decision framework
- 1. Explain: ask for the concept or one hint, not the finished submission.
- 2. Attempt: work the next step without looking at a generated result.
- 3. Retrieve: answer a practice question from memory before opening the explanation.
- 4. Check: compare reasoning, then repeat with a similar problem.
For VUST, use Study Practice or Exam-Format Practice for the question-and-hint workflow. The general solver has a separate task-solving role and can show a final answer.