your brain on accumulation of cognitive debt when using an ai assistant for essay writing task

How AI Essay Writing May Create Cognitive Debt

A great companion for students, writers, and professionals are AI tools like ChatGPT. Their use is rapidly growing, from brainstorming to full essay generation, large language models (LLMs) promise speed, fluency, and reduced effort. But a major question that arises beyond the convenience is : What will happen to the human brain if we offload thinking repeatedly to AI?

To answer these questions, Kosmyna et al. (2025), a study from researchers at MIT, investigated the neural consequences of LLM-assisted essay writing. It offers one of the most detailed answers so far. The researchers investigated whether frequent AI-assisted essay writing leads to what they call “cognitive debt,” using EEG brain recordings, linguistic analysis, and human evaluation. Their findings have been summarized in the article below.

“Cognitive debt” a gradual weakening of mental engagement that accumulates over time.

Click Here to download the Kosmyna et al. (2025) research paper.

What Is Cognitive Debt?

To understand the paper in detail we need to first know what Cognitive Debt is. It is a gradual weakening of mental engagement that accumulates over time. As in the case of technical debt in software development when you rely on shortcuts repeatedly, things work faster now but the underlying system weakens over time

In terms of cognition this simply means:

A reduced mental effort
Memory formation is weaker
Less or no ownership of ideas
Ability to retrieve or explain what you “produced” is reduced

In short, the MIT study wanted to test if AI tools merely assist thinking or subtly replace it.

How Researchers Tested AI’s Impact on the Brain

Participants and Study Design:

They recruited a total of 54 participants aged 18–39 from universities including MIT, Harvard, and Wellesley. Each one of them completed three essay-writing sessions, and a subset of 18 participants completed a fourth session months later.

Now the participants were divided into three groups:

The LLM Group - They wrote essays using ChatGPT only
Search Engine Group - They used Google and websites but no AI
The Brain-Only Group - They didn't use any tools

In the 4th session, the roles were reversed:

LLM to Brain - The LLM group who were long term AI users were forced to write without the help of AI
Brain to LLM - The Brain-only writers were allowed to use AI.

This role reversal was very important to detect long-term cognitive effects.

Measuring the Brain During Writing (With or Without AI)

This research study was not like most opinion-based studies, it used electroencephalography (EEG) to directly measure brain activity while participants wrote essays.

The EEG activity was analyzed across different frequency bands:

Alpha was used for monitoring attention and internal processing.
Beta for working memory and reasoning.
Theta & Delta measured integration and cognitive control.

The researchers also used the following:

NLP techniques - For analyzing essays using n-grams, named entities and semantic similarity.
They scored essays with the help of a specialized AI judge and human teachers.
After each session, interviewed participants.

All of these factors make the study one of the most comprehensive investigations of AI-assisted learning to date.

Less Thinking With More AI

This in simple terms says that brain connectivity scales down with the use of tools. The EEG results showed a clear hierarchy of cognitive engagement:

a. Brain-Only Group

This group had the most distributed and the strongest neural connectivity.
A higher engagement across memory, attention, and executive networks.

b. Search Engine Group

This group had moderate levels of engagement.
An active information synthesis and visual-executive processing.

c. LLM Group

Overall brain connectivity was the weakest in this case
A reduced alpha and beta coupling.
Rather than active reasoning, signs of cognitive offloading are seen.

In short, the more external help the participants received, the less their brains worked.

Session 4: Cognitive Debt Became Visible

Months later the most revealing results arrived, when the AI users had to think without the help of AI.

LLM-to-Brain: AI Users Had to Think Alone

The participants who were more dependent on ChatGPT for three sessions showed:

Lower neural connectivity persistently.
Poor or under-engagement in alpha and beta bands.
A difficulty in recalling what they had written previously
Continued reuse of AI kind of tools like vocabulary and paraphrasing.

Without AI, their brains did not come back to complete or full engagement. This is cognitive debt in action.

Brain-to-LLM: Independent Thinkers Used AI

The participants who were independent thinkers when given AI tools, showed:

An increased memory recall potential.
Strong activation levels in the prefrontal and occipito-parietal regions.
The engagement patterns were also similar to skilled search engine users

Instead of replacing their thinking, AI augmented it.

Linguistic Evidence: Essays Look the Same

In this case, natural language analysis revealed another important pattern that:

The essays written using LLMs showed a high within-group similarity.
Repetitive named entities, n-grams, and conceptual structures.
There is reduced semantic and stylistic diversity.

On the other hand, Brain-only essays consistently showed divergence from each other suggesting more personal framing and original reasoning.

Ownership, Memory, and the Illusion of Productivity with AI

Ownership - Did You Write This?

When the participants were asked to quote their own essays:

Around 83% of the LLM users failed
In the case of Search and Brain-only participants only ~11% of them struggled.

The self-reported ownership also followed the same trend:

The LLM group were among the lowest
The Brain-only group were highest

Interestingly, participants often described the AI-assisted essays as:

Technically good, well structured but emotionally distant or not truly “theirs”.

Human teachers also supported this view, describing many AI-assisted essays as polished but soulless.

Why AI Matters for Education

By default, the study does not argue that AI is useless or harmful. Instead, it highlights a crucial distinction: Learning requires cognitive effort and AI reduces the cognitive load.

LLMs make the task much easier by reducing the complexity. But if used as a replacement tool and not as an assistant or helping tool, they reduce

Deep memory encoding
The creative divergence
Learning gains that are long-term
Strategic reasoning

Consistent LLM users, over four months, underperformed across:

Originality in linguistics
Neural measures
Agency and ownership
Behavioral recall

Should We Stop Using AI for Writing?

Now the question is, should we stop using AI for writing. The answer is no. The study has a more nuanced takeaway:

Using AI after thinking is beneficial
If AI is used to thinking, it may prove to be costly.

The brain remains more engaged when learners first struggle, structure ideas, and then use AI to refine or expand. Cognitive debt begins to accumulate, when AI becomes the starting point.

The Bigger Picture: Your Brain on ChatGPT (AI)

This research has subtly reframed the AI debate proving that the real risk is not cheating it is atrophy, which means gradual weakening or loss caused by lack of use.

Just similar to how calculators changed how we teach math. AI will help us to reshape writing and learning. We will risk trading short-term productivity for long-term cognitive decline, if we don’t redesign how we use these tools.

In this scenario the question remains, “What happens to my brain if it always does?” and not “Can AI write this for me?”

**References: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872 (Kosmyna, N. et al.)**

Frequently Asked Questions (FAQs)

1. What is cognitive debt?

Gradual weakening of memory, mental engagement, and independent thinking caused by over dependence on external cognitive tools like AI.

2. Does this mean ChatGPT makes people less intelligent?

No, the study does not explicitly claim that AI reduces intelligence. It throws light on the fact that AI is used matters. Using AI as a refinement tool supports thinking and using it as a shortcut replaces thinking.

3. Why search engines perform better than ChatGPT?

The search engines make the users formulate queries, evaluate sources and integrate multiple viewpoints. All of these actively engage working memory and executive control networks, which the AI often bypasses.

4. Why couldn’t LLM users quote their own essays?

The LLM users' information was not deeply encoded into memory because much of the content was generated or heavily guided by AI. An effect known as cognitive offloading.

5. Is this effect temporary or long-term?

Yes, it suggests long-term effects. Prior LLM users showed reduced neural engagement, even after months and forced removal of AI tools.

6. Should students stop using AI for writing?

Students can use AI, but they should change how they use it:

Think first and try to write independently.
Make use of AI for structure, feedback, or revision.
Completely avoid AI-generated first drafts.

7. What does EEG connectivity actually measure?

EEG connectivity measures how strongly different brain regions communicate. Deeper cognitive processing and learning is indicated by stronger, distributed networks.

8. Does this apply outside education?

Yes, it is applicable for any domain that involves thinking, writing, planning, or reasoning may experience similar cognitive offloading effects with heavy AI reliance.

In short, AI does not just change what we write or how we write, it changes how our brains work while writing. If used wisely it can amplify learning. Otherwise, if used carelessly, it can quietly replace the very cognitive processes education is meant to build.