What if every shortcut cost you something you couldn't see?

Researchers at MIT put EEG caps on 54 people and watched what happened to their brains over 4 months of writing: with ChatGPT, with Google, and with nothing but their own minds.

What they found is worth 3 minutes of your attention.

Scroll to begin

54 participants. 3 groups. 4 sessions. 4 months.

🤖

LLM Group

Wrote essays using ChatGPT

🔍

Search Group

Wrote essays using Google

🧠

Brain-only Group

Wrote essays with no tools at all

Session 1Session 2Session 3Session 4
Their brain activity was recorded with EEG while they wrote. Their essays were analyzed with NLP. They were interviewed after every session.
In the 4th session, the researchers did something unexpected.

They flipped the groups.

More on that later.
MIT Media Lab, "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task," Kosmyna et al., June 2025. Preprint, not yet peer-reviewed. arxiv.org/abs/2506.08872

Your brain has muscles too. Here's what happens when AI does the lifting.

💡 Planning — "What do I want to say?"
🔍 Memory Recall — "What do I already know?"
✍️ Language — "How do I phrase this?"
🎯 Focus — "Stay on track."
🧩 Integration — "Connect the dots."

All cognitive muscles engaged.
Hard work, but you're getting stronger.

Now imagine outsourcing that work, session after session.
🧠 COGNITIVE DEBT STATEMENT
Account holder: You
Session 1 — AI drafted your intro −$5 🧠
Session 1 — AI structured your argument −$8 🧠
Session 2 — AI wrote most of your essay −$15 🧠
Session 2 — You edited AI's output −$3 🧠
Session 3 — AI handled everything −$25 🧠

TOTAL COGNITIVE DEBT: −$56 🧠
⚠️ Over 4 months and 4 sessions, LLM users consistently underperformed Brain-only participants at every level measured: neural, linguistic, and scoring.
💡 Meanwhile, Brain-only users built a SURPLUS of +$72 🧠 from regular cognitive exercise.
The researchers call this "cognitive debt."

It's not dramatic. It's not sudden. It's the quiet accumulation of small deferrals, moments where you could have thought, but didn't.

Each one is barely noticeable.
Together, over months, they reshape your neural landscape.

The Switch

LLM → Brain-only

After 3 sessions with ChatGPT, LLM users were asked to write alone for the first time.

Their cognitive muscles had been resting for months.

LLM group, writing alone Struggling
Brain-only group, using AI Supercharged

They still tried. Their delta and theta bands showed effort. But the higher-frequency processes, the ones that organize ideas, sustain reasoning, and drive self-expression, were significantly weaker.

They couldn't fully bounce back.

Brain-only → LLM

Meanwhile, Brain-only users were given ChatGPT for the first time.

Their well-trained brains didn't become passive. They activated more. Memory recall surged, prefrontal and occipito-parietal regions lit up. They used AI to amplify what they'd already built.

The tool didn't replace their thinking. It supercharged it.

"AI amplifies what you've already built.
It can't build the foundation for you."
The Pattern

It's not just one study.

Researchers are coming to similar conclusions.

🧠

MIT Media Lab

Kosmyna et al., 2025 · 54 participants · EEG brain monitoring · 4 months
"Brain connectivity systematically scaled down with the amount of external support."
Preprint, not yet peer-reviewed.
→ arxiv.org/abs/2506.08872
🎓

Wharton / UPenn

Bastani et al., 2024 · ~1,000 high school students · Math field experiment
"Students using unguarded ChatGPT performed 48% better WITH AI, but 17% WORSE on exams without it. A tutoring version with guardrails eliminated the harm."
Published in PNAS, June 2025.
→ pnas.org/doi/abs/10.1073/pnas.2422633122
🪞

Fernandes et al.

2026 · 698 participants · Two experiments · LSAT-style problems
"AI improved performance by 3 points, but users overestimated their own ability by 4 points. The more AI-literate, the worse the self-awareness."
Published in Computers in Human Behavior.
→ doi.org/10.1016/j.chb.2025.108779
🧩

Yurt & Kuşci

2026 · 422 university students · Structural equation modeling
"AI usage negatively predicted critical thinking (β = −0.37). Mediated by 'epistemic laziness,' the tendency to stop seeking deeper understanding when AI gives quick answers."
Published in Current Psychology.
→ doi.org/10.1007/s12144-025-08800-0
🌐

Gerlich

2025 · 666 participants · Mixed methods · Diverse age groups
"Significant negative correlation between frequent AI use and critical thinking. Younger users (17-25) most affected. Higher education provided some protective effect."
Published in Societies (MDPI).
→ doi.org/10.3390/soc15010006
← swipe to see more →
AI boosts your output today.
But it can erode your ability to produce output on your own tomorrow.

Unless you build the skill first.

This isn't about quitting AI.

It's about using it in the right order.

During AI use

Think first, then amplify

Write the first draft yourself. Use AI for the second pass. The MIT study showed that people who built skills first got MORE out of AI when they finally used it.

If you can't explain it without AI, you don't understand it

After using AI to help with something, close the tab and try to explain it to someone. If you can't, you consumed. You didn't learn.

Treat AI like a sparring partner, not a ghostwriter

Ask it to challenge your ideas, not generate them. Ask it to find flaws in your argument, not make the argument for you. The Wharton study found that AI tutors designed to give hints, not answers, completely eliminated the negative learning effects.

Building your baseline

Practice the hard thing regularly

The Brain-only group didn't have a genetic advantage. They just practiced thinking without tools, consistently. Three sessions of unassisted work built neural networks that LLM users couldn't match even after 4 months.

Notice the comfort

The most dangerous part of cognitive debt is that it feels good. Convenience feels like competence. If everything feels easy, ask yourself what muscle isn't being used.

Go Deeper