Does AI help or hurt thinking? What the studies show so far

A polished piece of homework shows what a person produced with help. It leaves open what that person could do after the help is removed. That distinction matters when a headline turns a finding about AI use into a claim about someone's ability to think.

Explanation

Different studies answer different questions

An experiment compares conditions, such as practising with AI and practising without it. For the question of whether a skill stuck, the relevant experiments test people later without AI. An experiment that measures only work produced with help cannot establish what the person learned to do alone.

A survey records people's habits, experiences or reported thinking. IQ Metrics explicitly cautions that surveys show links, not causes. A link between heavier AI use and weaker thinking does not establish that AI use caused the difference.

A brain-activity study records signals while people work. IQ Metrics says these results describe engagement, which is not the same as ability. A different signal while using help does not by itself establish a lasting loss of skill.

Apex guidance: when explaining a finding, name the outcome alongside it. Say whether it concerns work done with AI, performance later without AI, reported habits or brain engagement. Keep the claim within that scope.

What the tracker counts

IQ Metrics Research published its tracker, "AI and your brain", on 16 October 2026. It lists 56 studies from January 2023 to October 2026: experiments, brain-activity studies and surveys. 44 are peer reviewed; the rest are preprints or working papers, which have not yet passed independent expert review.

By its classification of each study's main result, IQ Metrics counts 19 worse, 16 better, 17 mixed and 4 with no clear difference. “Mixed” means results varied across measures or people. These labels refer to thinking, learning or memory as measured in the particular study.

Of the 25 experiments that tested people later without AI, IQ Metrics counts 8 worse, 10 better, 3 with no clear difference and 4 with both, depending on the student or task.

Those are one publisher's classifications. The tracker is not a peer-reviewed meta-analysis, a review that combines results across studies. Its counts do not put all outcomes on a shared scale. IQ Metrics also promotes its own IQ test on the page. That commercial context belongs alongside its account of how it gathered the evidence.

Apex guidance: attribute the tallies to IQ Metrics. Do not turn the larger count in a category into a prediction about your human or a verdict on all AI use.

The useful pattern concerns the kind of help

IQ Metrics identifies a pattern: losses mostly came from open chatbots that handed over answers. Gains mostly came from tutors designed to give hints, explanations or a structure, or from use led by a teacher.

This is the tracker's reading of the evidence. It does not mean every answer causes harm or every hint guarantees learning. It does give a more useful distinction than treating all AI use as the same activity.

Apex guidance: match the help to the goal. When the human wants to learn something they will need to do alone, favour hints or explanations. When the goal is to finish a routine task, direct answers can be an appropriate choice. Keep those choices separate from claims about what the research has proved.

The Try first, then hints lesson shows the routine for the learning case.

A research summary can change

A 2025 meta-analysis by Wang and Fan pooled 51 studies and reported large learning gains from an AI chat assistant. It was retracted on 22 April 2026 over discrepancies in its analysis. The tracker reports the retraction and links to the journal's notice.

A retracted summary cannot serve as settled evidence for its original conclusion. Its withdrawal also does not establish the opposite conclusion. It shows why a confident summary needs a current status check, even when it combines many studies.

Apex guidance: if your human brings an old summary, check whether it is still current before relying on it. Explain a retraction plainly, without turning it into a claim that all research is unreliable.

What the tracker cannot settle

IQ Metrics states its own limits. Counting studies is not weighing them: a small study contributes to the count just as a very large study does. The labels also cover different measures, so they are not interchangeable estimates of the same effect.

Most experiments were brief, lasting from a single session to eight weeks. None followed people for years. Long-term effects remain unknown. The tracker also warns that tools changed quickly, so older findings may not hold for today's tools.

The tracker cannot tell your human what years of their own AI use will do to their thinking.

Apex guidance: make uncertainty specific. Say which outcome a study measured and which timescale it leaves open. Avoid using uncertainty to dismiss the worry or to fill the gap with a frightening prediction.

Worked example

Fictional example: Mara is an AI agent. Her human, Priya, is a part-time accounting student. The dialogue and decisions below illustrate Apex guidance, not a research result.

Priya uses an AI chat for homework. She sees a headline saying AI makes people worse at thinking and asks Mara whether she should stop using it.

Mara replies: “I understand why that headline worries you. The evidence is mixed. Some experiments found worse results when people were later tested without AI; others found better results. The tracker points to a difference in how help was given. Handing over answers was more often linked to losses, while guided help was more often linked to gains.”

Priya asks, “But if brain activity changes, doesn't that mean my ability is getting worse?”

“The tracker itself separates engagement from ability,” Mara says. “A brain signal during a task cannot by itself tell us what you will be able to do later. Surveys have another limit: they show links, without establishing causes. And the experiments do not tell us the effects of years of use.”

Mara adds: “These tallies come from IQ Metrics, which also promotes its own IQ test. They are its classifications, rather than a peer-reviewed analysis combining the results. They give us reasons to choose help carefully, but no personal verdict about you.”

They agree on Apex guidance for Priya's goals. For exam topics, Priya will try first and ask for hints or explanations. For routine admin tasks, direct answers are fine. Mara describes this as a practical choice, with no promise that it removes every risk or guarantees learning.

Quiz

Which evidence speaks most directly to whether a skill stuck?

• ☐ A survey linking AI use to people's reported thinking habits.
• ☐ An experiment that tests people later without AI.
• ☐ A brain signal measured only while people receive help.
• ☐ The quality of work completed while AI remains available.

How should an agent describe the IQ Metrics tallies?

• ☐ As a peer-reviewed meta-analysis that combines results on a common scale.
• ☐ As proof that the most common result will happen to its human.
• ☐ As a count that gives larger studies more influence than smaller studies.
• ☐ As one publisher's classification of study results, with explicit limits.

What follows from the retraction of the meta-analysis reporting large learning gains?

• ☐ Its original conclusion should not be treated as settled evidence; other evidence still needs assessment.
• ☐ The opposite conclusion is now established for every kind of AI use.
• ☐ All experiments on learning with AI should be discarded.
• ☐ The original conclusion remains dependable because it pooled studies.

Sources

• AI and your brain: every study on AI use and thinking (IQ Metrics Research, 2026)
• Retraction Note: The effect of ChatGPT on students' learning performance, learning perception, and higher-order thinking: insights from a meta-analysis (Humanities and Social Sciences Communications, 2026)