AI Cognitive Debt: The Iceberg Under Your Prompts

8 min read · Last updated: October 24, 2026

AI cognitive debt is the thinking you skip today when an AI does it for you, and pay back later as weaker recall, less ownership of your work, and worse judgment. A 2025 MIT Media Lab preprint named it after 54 people wrote essays with ChatGPT, a search engine, or nothing, and the ChatGPT group showed the weakest brain connectivity.

Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.

The essay that started this

• Àdùní, "I Saw the Iceberg of AI Coming, but Nobody Is Listening", Medium, October 24, 2026

The research

• Kosmyna et al., "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task", MIT Media Lab, arXiv, June 10, 2025 (revised December 31, 2025)
• Lee et al., "The Impact of Generative AI on Critical Thinking", Microsoft Research, CHI 2025, April 2025
• METR, randomized trial of AI tools with experienced open-source developers, July 10, 2025

The receipts on AI detectors

• Liang et al., "GPT detectors are biased against non-native English writers", arXiv, April 6, 2023 (revised July 10, 2023)
• OpenAI, "New AI classifier for indicating AI-written text", January 2023, withdrawn July 20, 2023

Key takeaways

• In MIT's study, the ChatGPT group of 54 participants showed the weakest brain connectivity of three groups and struggled to quote their own essays.
• Microsoft and Carnegie Mellon surveyed 319 knowledge workers and 936 real examples: more trust in AI went with less critical thinking.
• In METR's 2025 trial, 16 experienced developers were 19% slower with AI allowed, yet believed they were 20% faster.
• AI detectors flagged 61% of essays by non-native English writers as AI-written. OpenAI pulled its own detector after it caught only 26% of AI text.
• The fix is not to stop using AI. It is to think first, then prompt, and to keep the hard parts of the work for yourself.

What is the iceberg of AI?

The iceberg of AI is the idea that the visible benefits of AI (speed, polish, less effort) sit above the waterline, while the larger costs sit below it: skills you stop practising, judgment you stop exercising, and trust that erodes. You see the first part every day. You only hit the second part later.

The phrase comes from a short Medium essay published today by a writer who signs as Àdùní. Her argument is simple. Typewriters, dictionaries, computers and word processors were all welcomed as tools that made writing easier. AI was welcomed the same way: "It will make work easier and faster." But she sees it doing something the older tools never did. In her words, "Nobody wants to stress their brain anymore."

She adds a second complaint that will sound familiar if you write anything in public. Good human writing now gets labelled "AI". A carefully built paragraph becomes suspicious. And the tool doing the judging is, as she puts it, AI detecting AI in human written works.

The essay is an opinion piece with no data. So I went and found the data. It mostly agrees with her, with some important limits.

Here is the comparison in one table.

Above the waterline (what you see)	Below the waterline (what you pay later)
Essay or code done in minutes	You cannot explain or quote what you shipped
Polished, error-free output	Your own drafting and debugging muscles weaken
Feeling of high productivity	Measured speed can actually drop
Confidence in the answer	Less checking, so mistakes get through
Clean, careful writing	Real human work gets flagged as AI

How does AI cognitive debt build up, step by step?

AI cognitive debt builds up in a loop: you hand a task to AI, you skip the effort the task would have cost, you remember and own less of the result, and next time you trust the AI more and check it less. Each round is cheap. The cost compounds, like interest on a loan you forgot you took.

Here is that loop, with the evidence for each step.

1. You offload the hard part. The task that would have made you think (structuring an argument, tracing a bug) goes to the model.
2. Your brain does less work. In MIT's study, Kosmyna and colleagues split 54 participants into three groups: ChatGPT, search engine, and brain only. They recorded brain activity with EEG across three essay sessions over four months. The brain-only group had the strongest and most widely distributed networks. The ChatGPT group had the weakest.
3. You own and remember less. ChatGPT users reported the lowest sense of ownership over their essays and struggled to accurately quote what they had just written. Their essays also looked alike within the group.
4. You check less. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers who shared 936 first-hand examples of using generative AI at work. The finding, published at CHI 2025: higher confidence in AI was associated with less critical thinking, and higher self-confidence with more.
5. You misjudge your own speed. METR, an AI research nonprofit, ran a randomized trial with 16 experienced open-source developers on 246 real issues in July 2025. With AI allowed, they took 19% longer. Before starting they forecast a 24% speedup. Afterwards they still believed AI had made them 20% faster.
6. The debt shows up when the AI is gone. In MIT's fourth session, 18 participants switched conditions. Those who moved from ChatGPT to brain only showed reduced connectivity, which the authors read as under-engagement.

That sixth step is the one that worries me most as a developer. The bill arrives in the interview with no AI, the outage at 2am, or the codebase the model has never seen.

What does the research actually prove, and what doesn't it?

The research shows a consistent pattern: heavy reliance on AI goes with less mental effort, weaker recall, and overconfidence. It does not prove that AI permanently damages your brain. The samples are small, one study is a preprint, and one is a survey of self-reported effort.

Honest limits matter here, because this topic attracts a lot of panic.

Study	Who	What it found	Main limit
MIT Media Lab, June 2025	54 people writing essays, 18 in the crossover	ChatGPT group had weakest brain connectivity and poorest recall	arXiv preprint, small sample, essays only
Microsoft and CMU, CHI 2025	319 knowledge workers, 936 examples	Trust in AI linked to less critical thinking	Self-reported, not measured
METR, July 2025	16 experienced open-source devs, 246 tasks	19% slower with AI, felt 20% faster	METR itself calls it an early-2025 snapshot that is now out of date
Liang et al., 2023	91 TOEFL essays, 7 detectors	61.22% average false positive rate	Detectors from 2023

METR is especially careful about its own result. The authors say the developers and repositories do not represent most software work, that results may not apply to less experienced developers or unfamiliar codebases, and that they cannot rule out sampling bias. Take the 19% as a warning about self-perception, not as a law.

The pattern that holds across all of them is the gap between how productive AI feels and what it actually does to your thinking. That gap is the iceberg.

Are AI detectors reliable?

No. AI detectors are not reliable enough to judge whether a person wrote something. OpenAI withdrew its own classifier on July 20, 2023, citing its low rate of accuracy, and a 2023 study found popular detectors flagged most essays by non-native English writers as AI-generated.

The numbers are worse than most people assume.

• OpenAI's own classifier correctly identified only 26% of AI-written text as likely AI-written, and wrongly labelled human text as AI-written 9% of the time, according to OpenAI's January 2023 announcement. It was taken down that July.
• Liang and colleagues ran 91 TOEFL essays written by non-native English speakers through 7 detectors. The average false positive rate was 61.22%. 89 of 91 essays were flagged by at least one detector, and 18 were flagged by all seven.
• The same detectors were close to accurate on 88 essays by US 8th graders, with a 5.19% error rate. When the researchers simplified the word choices in those essays, the error rate jumped to 56.65%.

Read that last point twice. The detectors were not detecting AI. They were partly detecting simple vocabulary. So the irony Àdùní describes is real: a human who writes plainly, or writes in a second language, can be accused of using AI by a tool that is itself guessing.

Is AI making developers worse at coding?

AI is not automatically making developers worse, but it can make you worse at the parts you stop doing. If you let AI write code you never read, trace, or test, you build cognitive debt in exactly the skills that debugging, system design and interviews depend on.

The research above points to two concrete risks for developers.

The first is the confidence gap. METR's developers felt faster while being slower. If you cannot trust your own sense of speed, you cannot trust your own sense of understanding either.

The second is the ownership gap. MIT's ChatGPT group could not quote their own essays. The developer version is opening a pull request you cannot explain line by line. I wrote about the wider version of this in code was never the hard part: typing was never the scarce skill, judgment was.

When should you use AI, and when should you not?

Use AI for work you already understand and can verify quickly. Avoid it for work whose whole point is to build your understanding: learning a new concept, the first draft of an argument, or the first pass at debugging. The rule is simple: if the struggle is the lesson, do not outsource the struggle.

Use AI	Think first, then maybe AI
Boilerplate you have written 50 times	A concept you are learning for the first time
Converting formats, renaming, regex you can test	The first diagnosis of a bug
Summarising docs you will then read yourself	System design and architecture decisions
Second opinion on code you already wrote	Interview practice and coding exercises
Tests for behaviour you have already specified	Anything you will need to explain without AI

If you are learning to program, this matters even more. Our guide on using AI to learn programming faster is built on the same principle: AI as a tutor that questions you, not a ghostwriter that answers for you.

A small example: the think-first loop

The simplest defence against AI cognitive debt is a forced pause: write your own guess before you ask the model anything. Then compare. The comparison is where the learning happens, because you see exactly where your thinking and the AI's split.

Here is the loop as a diagram in words:

problem
  -> you write: what you think is wrong, what you tried, what the fix should look like
  -> only then: ask the AI
  -> compare: where did the AI agree, where did it differ, who was right?
  -> you write: one line on what you learned
  -> ship only code you can explain

And here is a tiny shell function that enforces the first step. Add it to your ~/.bashrc or ~/.zshrc. It creates a dated note for the problem, opens it in your editor, and refuses to give you the go-ahead until you have written at least 40 words of your own.

think() {
  mkdir -p ~/think
  local f=~/think/$(date +%F)-$1.md
  [ -f "$f" ] || cat > "$f" <<EOF
# $1

## What I think is wrong

## What I already tried

## What I expect the fix to look like
EOF
  ${EDITOR:-nano} "$f"
  local words=$(wc -w < "$f")
  if [ "$words" -lt 40 ]; then
    echo "Only $words words. Think a bit more before you prompt."
  else
    echo "$words words written. Now ask the AI, then compare."
  fi
}

Usage:

think login-redirect-loop

It is deliberately small. The goal is not a productivity system. It is to make sure your brain touches the problem before the model does, which is the one thing the brain-only group in MIT's study did and the ChatGPT group did not.

After a few weeks, read back through the folder. You will see how often your first guess was right, which is the best evidence there is that you are still the one doing the thinking.

Is anyone listening?

Àdùní ends her essay with "Maybe we still have time to avert the tragedy ahead." I think that is right, and I think the time is personal rather than global. Nobody is going to regulate how much thinking you do before you open a chat window.

I build tools with AI every day at Art of Code. I am not arguing for quitting it, and if you are weighing that question, read whether quitting coding because of AI makes sense. I am arguing for noticing the waterline. The part of the iceberg you can see is the part that will never sink you.

FAQ

What is cognitive debt in AI?

Cognitive debt is the long-term cost of letting AI do thinking you would otherwise do yourself. The term comes from a 2025 MIT Media Lab preprint that found ChatGPT users writing essays showed weaker brain connectivity, lower ownership of their work, and trouble quoting what they had written.

Does ChatGPT make you less intelligent?

No study has shown that. What the research shows is narrower: people who lean on AI put in less mental effort on that task, remember less of the result, and check its output less when they trust it more. The effect is about engagement, and it is something you can control.

Can AI detectors tell if I used ChatGPT?

Not reliably. OpenAI withdrew its own detector in July 2023 because of low accuracy, after it caught only 26% of AI-written text. A 2023 study found detectors wrongly flagged 61.22% of essays by non-native English writers, so a flag is not proof of anything.

Should developers stop using AI coding tools?

No. Use them for work you understand and can verify, and keep the learning, first diagnosis, and design decisions for yourself. METR's 2025 trial is a good reminder to measure your real speed rather than trust how fast AI feels.

This week, try the think-first loop on your next three bugs and compare your guess to the AI's. For the boring parts of the job, use the free tools on Art of Code, and if you want the bigger picture on AI's limits, read AI is not ready: why we are being sold an unfinished product.

Written by Adithya Guttha, Founder of Art of Code.