AI & Career
Cognitive Debt: What the MIT Study Reveals About AI That Thinks for You
AI & CareerNov 12, 2026

Cognitive Debt: What the MIT Study Reveals About AI That Thinks for You

Understand the concept of cognitive debt and what MIT research reveals about the risks of letting artificial intelligence think for you.

BlenseReading: 7 min

A text written with AI about the cost of not thinking

This article about cognitive debt was built with the help of artificial intelligence: I chose the topic and the direction, debated and reviewed every line with the model, and now it's up to you to judge whether I paid that bill or fell into the machine's trap.

A feeling you've probably had

You ask AI to write an email, a paper or a piece of code. It's ready in seconds, looks good, and you hand it in. The next day, someone asks: "Why did you write it this way?". And you realize you couldn't explain.

The text has your name on it, but it barely passed through you.

This is the feeling researchers call cognitive debt.

The idea in one picture

Think of a credit card. You buy something today and take it home right away. The bill doesn't disappear: it arrives later, with interest.

With AI, the logic is similar. When you let the tool do the mental effort, you save time now. The "interest" is paid later: remembering less of what you did, feeling less that the idea is yours and struggling more when you need to think on your own.

Important: this is a metaphor. No one has measured an actual "bill" in the brain. It's a way of naming a suspicion that makes sense.

The experiment that named the idea

In 2025, researchers at MIT, a university in the United States, ran a simple test. They gathered 54 people and asked them to write essays in three groups:

  • one group using ChatGPT;
  • one group using a search engine, such as Google;
  • one group using only their own heads.

Participants wore sensors on their heads to measure how different regions of the brain "talked" to each other while writing.

What they saw, in short:

  1. A less connected brain. Those who wrote without tools showed more coordinated activity. Those who used ChatGPT showed less. The search engine group fell in between.
  2. Weaker memory of their own text. Right after writing, many ChatGPT users struggled to repeat a sentence from the essay they had just handed in.
  3. Less sense of authorship. The AI group felt less that the text was theirs, and their essays ended up looking alike.

In an extra session, some participants switched groups. Those who came from AI and started writing on their own seemed "slow to get back into the rhythm". Meanwhile, those who had been writing on their own and switched to AI showed more activity, as if they were fitting the machine's suggestion into the idea they already had.

The most interesting clue is this: order seems to matter. Thinking first and using AI afterwards is not the same as starting with AI.

Before drawing conclusions: what the study does NOT prove

The headline that went around was "AI makes us dumb". The authors themselves asked people not to say that, and rightly so:

  • It's a small study. Only 54 people, and just 18 did the switching session.
  • It hasn't yet been reviewed by other scientists, as far as I could verify. It was released as a "preprint", a preliminary version.
  • It was a single task: a 20-minute essay, with a single AI model.
  • It measured a few months, not years. It doesn't prove that AI causes lasting damage to the brain.
  • Other researchers criticized the method and the data analysis.

So the honest reading is: an interesting clue, not a verdict.

Other studies point in the same direction

MIT is not alone. Two studies help explain the pattern:

High school students. In an experiment with about a thousand students, those who practiced with a regular ChatGPT did very well during the exercises. But on the unassisted exam, they scored about 17% below those who never used it. Meanwhile, a group that used an AI tutor that gave hints instead of ready-made answers tied with the control group on the exam. The difference between the two bots was how they were designed to teach.

Programmers learning something new. In an Anthropic test with 52 engineers, those who coded with AI help scored an average of 50% on a quiz afterwards, versus 67% for those who coded on their own. Within the AI group, those who just had the machine do everything did worse. Those who asked questions to understand, or asked for the code and then studied it, did better.

The common thread: having AI is not what gets in the way. It's handing it precisely the part of the reasoning where learning happens.

Investment or debt?

Back to the credit card: it's neither good nor bad. Used for something that pays for itself, it's an investment. Used to dodge a bill, it becomes debt. AI is similar:

  • Investment: you use AI to think better. It explains, questions, gives hints, and you do the work. The effort stays with you, and so does the learning.
  • Debt: you use AI so you don't have to think. The result comes fast, but the effort didn't disappear; it was only postponed.

The same tool can be both. The difference is in how you use it.

And in the world of software development?

Imagine an engineer using AI to rush out a feature on a Friday, shortly before the end of the workday. The code works, the tests pass and the task is closed. Now imagine ten other engineers doing the same, each shipping parts of the system without fully understanding the decisions behind the generated code.

The problem may not show up that day. But what happens when a bug appears in production? Whoever has to investigate will need to understand not only what the code does, but also why it was written that way. If no one can answer, the time saved during implementation may be lost — or even multiplied — during investigation and fixing.

In software development, this risk is often discussed as a gap in understanding (comprehension debt): the system keeps working, but the team accumulates code whose logic it doesn't understand well enough to maintain, modify or troubleshoot safely.

The same caveat applies here: this is a metaphor used in management, backed by developer accounts, not a lab result.

What to do in practice

Five simple habits, based on the clues above:

  1. Think first, use AI afterwards. Sketch your idea, plan or logic before opening the tool.
  2. Ask for explanations, not just answers. "Why does this work?" is worth more than "do this for me".
  3. Ask for hints when you're learning. Leave the final answer to yourself.
  4. Review for real. Reread the result and check it, instead of just accepting it.
  5. Do the close-the-window test. Close the window and explain the text or the code in your own words. If you can't, the debt is growing.

Takeaways

Cognitive debt is a good name for a real feeling: the text is done, but it isn't ours; the system works, but no one knows why.

We have no proof that this causes lasting damage. We have clues, from different places, that using AI in place of mental effort delivers fast results and weaker understanding. And that the way you use it makes a difference.

Before each task, it's worth asking whether what you're doing with AI is an investment or a debt. In practice, it comes down to one simple question: what do I want to keep in my head from this, beyond the ready-made answer?

Sources: