AI and thinking

Two claims dominate discussions about AI in schools: that it frees students for more demanding intellectual work, and that it weakens their ability to think independently.

Neither is an adequate basis for school policy. Research has found benefits and harms under particular conditions. Schools need to examine those conditions and distinguish a better finished product from stronger understanding.

Where assistance may help

AI offers several possibilities worth testing in instruction.

A starting point. A question, example, or suggested approach may help a student begin. The appropriate support depends on the task. If constructing an argument is the learning objective, supplying a complete argument may remove the work the student needs to practice.

Feedback during revision. A tool can respond while a student is still working. Its usefulness depends on whether the comments are accurate, relevant to the assignment, and understood by the student. Accepting suggested changes is different from learning how to make them.

Help with incidental tasks. Assistance with formatting or organizing materials may leave more attention for the learning objective. Teachers should decide which steps are incidental: selecting a mathematical method and carrying it out can both require knowledge students need to develop.

Additional explanations and practice. Students can request another example or explanation when they need it. That creates an opportunity for support, provided the response is sound and the student does something with it beyond reading.

What studies have found

A randomized study involving nearly 1,000 high school mathematics students in Turkey illustrates the difference between assisted performance and learning. Students using a version of GPT-4 that would answer their questions directly performed better during supported practice but worse on a subsequent unaided assessment than students without AI access. Most of them asked it for the answer, which the researchers identify as the cause. A second version, designed to give teacher-written hints without the answer, largely avoided that harm, though it did not produce a significant improvement on the unaided assessment. Bastani and colleagues, PNAS

A separate study had 194 students in a Harvard undergraduate physics course take one lesson at home with a purpose-built AI tutor and another in class, in random order. Immediate learning gains were greater after the AI lesson. The tutor was deliberately designed around instructional principles; the result does not establish that unrestricted chatbot use would have the same effect, and the two lessons differed in setting as well as method. Kestin and colleagues, Scientific Reports

Together, these findings support a practical conclusion: the design of the assistance matters. Access to AI alone tells us little about what students will learn.

What schools should watch for

A student may produce an accurate answer without being able to explain or reproduce the reasoning. That gap deserves attention whether the assistance came from AI, a worked solution, or another person.

When evaluating AI use, ask what the student practiced. Did they select a method, test an interpretation, retrieve relevant knowledge, or revise an argument? Or did the tool perform those steps while the student checked that the result looked plausible?

Difficulty is not automatically productive. Confusing instructions and inaccessible materials can obstruct learning. The aim is to preserve the intellectual work that develops understanding while providing enough support for students to attempt it successfully.

Time saved is therefore an incomplete measure. Schools also need evidence of what students can subsequently explain and do.

Where the evidence remains limited

These studies address particular tools, subjects, students, and short-term outcomes. They do not settle the effects of years of routine AI use on retention, independence, or the ability to apply knowledge in unfamiliar situations.

Schools should examine research claims at that level of detail. Who participated? What assistance did they receive? What was the comparison? Were students assessed with or without the tool, and how long afterward?

A study can provide useful evidence without establishing that the same result will occur across a district.

What this means for assignment design

Identify the thinking an assignment is intended to develop before deciding where AI assistance belongs.

In research, that may include finding and evaluating sources. In mathematics, it may include choosing a method and explaining why it works. In writing, it may include developing an argument as well as expressing it clearly.

Then specify the assistance students may use and gather evidence of the learning that remains their responsibility. A brief independent task, an explanation of a revision, or a later application of the same concept can help teachers assess whether assistance supported understanding.

Our guide to what students should disclose about AI use explains how students can account for that assistance.

The judgment students need

Students should learn to recognize when a tool helps them make progress and when it is doing work they need to practice. They should also be able to check its contribution and explain the reasoning behind their final work.

Teach those decisions through examples, guided practice, and opportunities to work independently. Our guide to AI literacy describes how that instruction can develop across the curriculum.