One of the most useful distinctions in learning research is easy to miss in an AI-tutor demo. Remembering a lesson and using it on a new problem are different achievements.

A small randomized study published this summer makes the difference unusually clear. Researchers assigned 34 undergraduates to learn histogram equalization, a technique from digital image processing, through either a dialogue with ChatGPT or a human instructor.

The two groups performed comparably on a retention test. On the transfer test, the students taught through ChatGPT did significantly worse.

Thirty-four students learning one technical topic is nowhere near enough evidence for “human teachers beat AI.” It is enough to ask a better question about what an AI tutor should be expected to produce.

Imagine that a student has learned a worked method for changing the distribution of brightness values in an image. A retention question can ask for the idea again, or for steps close to the ones just practiced. A transfer question changes the surface of the problem. The learner has to recognize that the same knowledge is relevant, decide how to use it, and adapt it without being handed the original path.

The second task is often what we mean when we say somebody can actually use what they learned.

The EEG result is interesting, but not the headline

The researchers also recorded EEG while the students learned. The ChatGPT group showed broadly higher theta power, and some regional gamma differences survived the authors' later corrections. Those findings could sound impressive if translated loosely into “AI made the brain work harder.”

The paper itself is more restrained. The pattern is consistent with greater visually driven cognitive effort during a text-based interaction, not with deeper learning simply because more neural activity appeared. Some post-dialogue EEG results also came from a short resting window and deserve caution.

Behavior gives the cleaner result here. The students could retain the material about as well, yet they had more trouble carrying it into a new problem.

That separation matters well beyond AI. A student can memorize a formula without learning when to choose it. A chess player can remember a tactical pattern from a lesson and fail to notice the same structure when the pieces are rearranged. A programmer can follow an example and still freeze when the specification changes.

A tutor can therefore improve one target while leaving another underdeveloped.

Other AI tutors have done much better

This study should not be turned into a verdict on AI tutoring because “AI tutor” is not a single intervention.

A 2025 randomized study in Scientific Reports tested a purpose-built AI tutor in an undergraduate course and found that students learned more in less time than students in an in-class active-learning condition. The tutor had been deliberately designed around research-based pedagogical practices rather than simply opening a general chatbot and beginning a conversation. Students also reported strong engagement and motivation.

Other experiments have produced still different results. One 2025 study of unrestricted ChatGPT use as a study aid found worse performance on a surprise delayed test 45 days later. Another set of experiments on AI summaries and reflective prompts around educational videos found no clear retention or transfer advantage in that setup.

These findings do not combine into a neat ranking. They point toward the part of the problem that is actually useful: system design changes the cognitive work the learner has to do.

A tutor may explain more or less, ask the learner to generate an answer before helping, provide hints instead of solutions, require retrieval, vary the examples, or deliberately test the same principle under new surface conditions. Those choices are part of the learning intervention.

A simple transfer check for AI-assisted study

If you use an AI tutor, the easiest way to learn something from this research is not to stop using it. Add an independent transfer check.

After the explanation is over, try a problem that is meaningfully different from the example you just studied. Do it without first asking the AI which method applies. If you cannot decide how to begin, that is information about what remains to be learned.

For factual material, transfer may be less important than accurate recall. For a procedure you expect to use in unfamiliar situations, it can be central. The test should match the capability you care about.

That is also why a tutor that feels excellent can be hard to evaluate while you are using it. Helpful dialogue can make the current problem move smoothly. The more revealing moment may come afterward, when the scaffolding is gone and the next problem does not announce which lesson it belongs to.

Sources

Ulix app

Guten

Guten is a free Android reader for public-domain classics, with tools for highlighting and notes. For material you actually need to learn, the important step still happens after capture: trying to explain, retrieve, or use it without the text doing the work for you.