Study shows AI-assisted homework raises scores 18% but cuts exam performance 20% without tools

The study tracked 27,000 students aged 12 to 18 in China, where 80% reported using AI tools like Doubao and DeepSeek for homework. The remaining 20% served as a control group. Over six months, AI users saw an 18% improvement in homework scores across subjects. However, when tested without AI, their exam scores dropped 20% below those of non-users. This suggests AI assistance may inflate performance metrics without improving underlying comprehension or retention.

The findings align with earlier research from the University of Pennsylvania, where students using AI for math practice outperformed peers in short-term exercises but lost that advantage in closed-book tests. The pattern indicates AI tools may encourage surface-level engagement, such as copying generated answers, rather than deeper learning. For engineers, this highlights a critical limitation: AI can optimize for immediate output but may fail to reinforce durable skills.

The study’s design does not isolate specific AI features or usage patterns that drive the performance gap. For example, it’s unclear whether students used AI for problem-solving, content generation, or rote memorization. Without granular data, it’s difficult to determine whether the issue lies with the tools themselves or how students interact with them. This ambiguity complicates efforts to mitigate the negative effects while preserving AI’s potential benefits.

The broader context shows rapid AI adoption among students, with surveys reporting 80-94% usage rates in wealthy countries. Teachers have noted an uptick in formulaic, AI-generated submissions, but robust evidence on learning outcomes has been scarce until now. The study’s results challenge the assumption that AI-assisted learning translates to long-term gains, which could influence how educational institutions integrate these tools.

For engineers, the study underscores the need to design AI tools that encourage active learning rather than passive reliance. Features like step-by-step explanations, interactive problem-solving, or adaptive feedback could help bridge the gap between assisted and unassisted performance. However, the trade-off between convenience and skill development remains unresolved, and further research is needed to identify best practices.