Research has found that groups who studied with AI tools scored as much as 20% lower than others when tested under conditions that excluded AI.

The study divided participants by whether they used AI during the learning phase, then had both groups sit examinations without AI.

The idea of desirable difficulty

Cognitive psychology has a concept called desirable difficulty: when learning is too smooth, immediate comprehension comes easily but retention suffers. A body of research holds that struggling to retrieve an answer oneself — retrieval practice — is central to forming long-term memory.

AI tools by definition remove that difficulty. Getting an explanation the moment one is stuck skips the retrieval attempt entirely.

Limits of interpretation

It is nonetheless hard to reduce this to 'AI harms learning.' Because the assessment excluded AI, what was measured is the ability to perform without it. If real working conditions permit AI, which abilities to cultivate is a separate question.

How AI is used matters as well. Taking an answer directly and drafting one's own answer before asking for review may have different learning effects.

Practical implications

Educators are discussing setting different tool-use rules for learning and assessment phases. This is especially relevant for students at Chinese schools studying in multilingual settings: leaning on translation tools solves the immediate assignment but can slow the accumulation of language ability itself.

This paper recommends no particular study method. The practical lesson is to draw a conscious line between when to use tools and when not to.