AI assistance is not the same as learning

Getting a better answer with AI and learning how to produce, explain, or transfer that answer are different outcomes.

One field experiment in high-school mathematics found exactly this separation: students using unrestricted GPT-4 performed better on assisted practice, yet later performed worse without the tool than students who never had access. A tutor version designed to give structured help rather than simply provide solutions largely removed that harm.

That does not justify an anti-AI conclusion. A separate randomized study in introductory university physics found substantially stronger short-term learning gains with a carefully designed AI tutor than with the comparison active-learning lessons. The tutor was deliberately scaffolded around pedagogical practices, expert content, self-pacing, and targeted feedback.

The more defensible conclusion is therefore about design and measurement:

• assisted performance is not sufficient evidence of independent capability;
• AI can support genuine learning when the interaction is designed for learning rather than only task completion;
• the relevant test is what the learner can understand, retain, transfer, and do when assistance changes or disappears.

This connects directly to Learning is not compliance: a system that makes output easier is not automatically a system that expands capability.

Explore connections ↓