What is cognitive debt in AI learning research?

MIT Media Lab's preprint "Your Brain on ChatGPT" used the phrase for a pattern in adult essay writers. ChatGPT users showed lower EEG connectivity while writing, and some of that reduction lingered when they later wrote without AI. It is an early signal, not settled science.

What is cognitive surrender?

Wharton researchers Steven Shaw and Gideon Nave describe people adopting an AI answer as their own without noticing the handoff. In their experiments, accuracy rose when AI was right and fell below baseline when AI was wrong.

How can AI help without doing the student's thinking?

Teacher-facing AI diagnoses misconceptions after assessments. Plan, Teach, Diagnose, Verify. The student still writes, solves, and revises. Students never talk to the AI.

Cognitive Debt, Cognitive Surrender, and the Case for Letting Students Do the Work

Q: Should schools ban AI because of these studies?

No. Both studies are early, and the authors say so. They point to a design principle, not a prohibition. Keep the thinking with the learner. Put AI on the evidence the teacher reads.

Two phrases have taken over the AI-and-learning conversation.

Cognitive debt. From MIT Media Lab's "Your Brain on ChatGPT" study.

Cognitive surrender. From Wharton researchers Steven Shaw and Gideon Nave.

Both describe a version of the same thing: when a system does the thinking, the person stops doing it, and often does not notice the handoff.

For schools, that is the whole ballgame. Neither finding is settled science. Both are useful warnings.

What the MIT study found

Researchers led by Nataliya Kosmyna had 54 participants, ages 18 to 39, write essays under three conditions: with ChatGPT, with a search engine, or with no tools. They measured brain activity with EEG.

The ChatGPT group showed the lowest brain connectivity while writing. In a follow-up phase where former ChatGPT users wrote without AI, their connectivity remained lower than that of participants who had never used it. Human evaluators, blind to the conditions, described the AI-assisted essays as formulaic, with limited vocabulary variation. The researchers' word was "soulless."

Important caveats, from the authors themselves: the study is a preprint, not yet peer-reviewed. The sample is small and adult. Kosmyna has said plainly, "We didn't find any brain rot," and that the study did not measure IQ. She has called for more research across age groups and tasks.

So: a real signal, not a verdict.

What the Wharton study found

Shaw and Nave ran multiple experiments with more than 1,300 participants and nearly 10,000 individual trials. The paper, Thinking - Fast, Slow, and Artificial, is on SSRN.

The headline result is symmetrical and uncomfortable:

When AI gave a correct answer, participants' accuracy rose about 25 percentage points.
When AI gave a wrong answer, accuracy fell about 15 percentage points below where participants started.

People did not use AI as a second opinion. They adopted its answer as their own, "without recognizing that a transfer took place." That is the definition of cognitive surrender, and what separates it from ordinary offloading, where a person still owns the decision.

Treat this as a research framing, not a closed case. The paper is on SSRN; reporting appeared in 2026.

Why this is an education problem before it is a workplace problem

An adult who surrenders judgment to AI has, at least, judgment to surrender.

A ten-year-old learning to add fractions does not yet. If the answer arrives before the struggle, the skill never forms. There is nothing to offload because nothing was built.

This is the concern that New York City named when it paused student-facing generative AI for grades 2-K through 8, and that Norway's prime minister named when he said AI lets children skip important steps in their education.

It is also why the output of student work is such a poor signal. An essay can be fluent and empty. A worksheet can be correct and unlearned.

The practical response is not "no AI"

Both studies point to a design principle, not a prohibition.

Keep the thinking with the learner. Put the AI somewhere else.

Where? On the evidence.

A teacher facing 28 students cannot personally trace every wrong answer to its root cause. That is a place where a system can help without touching the student's cognition at all:

Which students shared the same misconception today?
Is the fraction error really a place-value gap?
Did the reteach change anything?
Who is ready to move faster?

The student still writes the essay, solves the problem, and revises. The teacher gets sharper information about what to do next.

How we built for this

Omnix360 is instructional infrastructure that runs behind the teacher. Students never interact with the AI. The loop is Plan → Teach → Diagnose → Verify, and every teacher override makes the diagnosis more accurate over time.

The cognitive load that MIT and Wharton warn about is the student's to carry. The information load, which is what actually burns teachers out, is the part worth automating.

See AI in the classroom, teacher-facing and the AI performance cliff.

The takeaway

Cognitive debt and cognitive surrender are early findings, and the researchers say so. But both point the same direction.

The value of a school is not in the answers it produces. It is in the thinking it builds.

AI that does the thinking for a child is a liability. AI that shows a teacher where the thinking broke down is an asset. The difference is entirely in the design.

See Teacher-Facing AI in Action.

Sources

MIT Media Lab: Your Brain on ChatGPT (preprint) (2025; 54 adult participants; authors caution it is not peer-reviewed)
Shaw and Nave, Thinking - Fast, Slow, and Artificial (SSRN)
New York City Mayor's Office: generative AI moratorium (September 2, 2026)
Reuters: Norway imposes near ban on AI in elementary school (June 19, 2026)