How ChatGPT makes you dumb.
But only if you make this mistake (& how to avoid it).
I use ChatGPT every day.
And if you read this newsletter, you probably do too.
Well… in this case…
The MIT said we’re dumb.
Generate it & forget it.
MIT asked 54 students to draft SAT-style essays to track how method affects recall, writing quality, and real-time brain activity.
☑ 18 students wrote with ChatGPT only,
☑ 18 students wrote with Google search, and
☑ 18 students wrote with nothing but their own brains (gasp).
4 minutes after writing an essay with ChatGPT, 15 of 18 “ChatGPT-only” participants failed recalling a single sentence (83%).
This is much more than the others students (11%) not using ChatGPT.
MIT researchers call the gap “cognitive debt”: when generative text appears on-screen, users skim, accept, and move on.
Worse: brain-only writers felt “total ownership” of their essays, whereas ChatGPT users said it “felt half-mine at best” .
And that’s not it. Even their brains looked different during the test.
ChatGPT makes your brain go numb.
What they measured: While students wrote, the researchers tracked how many brain areas were talking to each other at the same time (EEG). Think of it as counting live phone lines inside your head.
What they found: Using ChatGPT cut the brain-to-brain traffic by roughly half.
☑ Brain-only writers: about 80 active lines.
☑ ChatGPT writers: about 40 active lines.
How to interpret it: Fewer active lines means your own mental circuits aren’t coordinating as much. The tool is supplying ideas, structure, and wording for you.
Why it matters for you: If you let the model do most of the thinking, you practice less thinking yourself. So you numb yourself with AI.
But this study is far from being perfect. It has blind spots.
4 blind spots.
Task diversity. All writing were SAT-style essays. Would coding, poetry, or slide-design tasks produce the same trade-offs?
Long-run. They did it over 4 months, but it’s still short of the years we will be using AI to form—or fix—habits. For better or worse.
Realism. Participants wrote essays in a lab, wearing 32-channel caps; and one single tab allowed. No coffee. No Slack. No Whatsapp. No Instagram. This is not the real world.
AI skills. The paper controls for age and gender, not for prior writing skill or knowing how to AI. And that makes all of the difference. You & I both know it since we’re writing (me) and reading (you) ‘How to AI’.
The MIT study sure is a wake-up call, but not a death sentence.
They prove that passive AI usage might damages recall & reduces cognitive ability.
But at the end, they also hint that active usage patterns, where you stay in the loop, is a totally different story.
In the next section of our guide, I will explain how AI makes people more creative & productive (= smarter?) when we start looking deeper.
Then, I will pivot from “ChatGPT makes me dumb” to “How to (actually) use ChatGPT to be smarter”, mapping precisely how to keep our human brain in charge.
Let’s start with why, and then how.
AI is a means, not the end. And it pays off.
MIT’s own “Delay-Then-Augment” playbook.
In the 4th writing session, the Brain-only writers were finally given ChatGPT, and their brains did the opposite of “shut down.”
Alpha-, beta-, theta-, and delta-band networks surged above every earlier session (without AI), and independent graders awarded their essays the highest overall scores.
The researchers sum it up:
Rewriting an essay using AI after prior AI-free writing engaged more extensive brain-network interactions.
Translation: do the hard thinking first, invite AI in second.
That order keeps your memory circuits active while letting the AI polish, fact-check, or expand ideas you already own.
Practical takeaway. Write a quick outline or even a rough first draft on your own.
Only then open ChatGPT to challenge your structure, add counter-arguments, or surface examples you missed. You stay in the cognitive driver’s seat; the model becomes a booster, not the steering wheel.
Humans who pair grit + GPT win big.
AI is nice, but does it pay the bills? Yes.
A field experiment with 758 Boston Consulting Group employees found that consultants who used GPT-4 finished tasks 25% faster and produced work that judges rated 40% higher in quality than a control group typing solo.
Half of consultant’s work is 1) research 2) making slides. So coupling any AI deep research like ChatGPT/ Perplexity to making slides with gamma.app sure makes one better at its craft.
Developers show the same pattern. GitHub’s multistudy analysis reports that Copilot users cut coding time by 55% while improving readability and error rates. (resources.github .com)
GitHub’s telemetry shows time-to-merge cut in half and developers reporting higher job satisfaction because the “grunt work” moved to the bot. (github .blog)
Even at the bleeding edge of competitive coding, “GPT-o3” recently got a 2700 Codeforces rating. And that’s waaaaaay better than most people. (codeforces .com)
Creative work. Controlled trials in Science Advances find that stories drafted with AI assistance are judged more creative and engaging, especially for less-experienced writers. They still require a human to curate and vary ideas to avoid sameness. (science .org)
Bottom line. None of these top performers copy-paste and call it a day.
They iterate, reframe, and blend the AI output with their own expertise.
And they end up being both faster and better.
Now that we know why, time to master how.
How to AI. To get smarter, not dumber.
6 concrete workflows that preserve your brain while using ChatGPT multiply your reach. Each one maps to an insight from the studies & my experience.
Workflow 1. “Draft, Then Ask”
Goal: understand before the model begins suggesting phrasing or arguments.
Why it works: participants who wrote first and only added GPT later showed the largest spike in whole-brain connectivity and the best essay scores.
Step-by-step:
10-minute solo draft.
Ignore spelling; just pour ideas onto the page.
Time yourself. You will overcome the writer’s block.
Gap scan.
Mark areas that feel thin, repetitive, or weakly evidenced.
Targeted prompts.
Open ChatGPT now and pour in the context. If you write a paper, share it (as a PDF).
Then, finally ask specific questions: “Give me three counter-arguments to Point B,” or “Suggest two historical examples for my thesis.”
Curate, don’t copy.
Paste only what genuinely strengthens your narrative.
Rewrite it in your own tone if necessary. Add your spice.
Or train ChatGPT on your tone (by giving it a snippet).
Workflow 2. “Socratic Chat”
(Ideal for the ones who are afraid of confirmation bias)
I made an entire guide on How Socrates prompts ChatGPT.
Goal: turn ChatGPT into a relentless critic that sharpens your reasoning rather than writing for you.
Why it works: challenging the model (instead of accepting its first answer) keeps the fronto-parietal “executive control” loops active. You stay in the loop.
Step-by-step:
State your thesis to GPT.
“I argue that universal basic income boosts entrepreneurship.”
Ask GPT to destroy it.
“List the three strongest objections, with evidence.”
Defend / revise.
For each objection, either rebut or concede and adjust your claim.
Ask GPT to rate the new version.
“On a scale 1–10, how internally consistent is this draft?
Where is it weakest?”
Repeat until GPT’s score ≥ 8
and you can verbalize every point without the prompt visible.
Workflow 3. “AI Pop-Quiz Loop”
(Ideal for the students, or the ones who need to learn something)
Goal: convert ChatGPT into a just-in-time memory trainer so new information sticks.
Why it works: the study’s biggest red flag was 83 % of ChatGPT users failing to recall their own sentence . We instead use ChatGPT as a studying buddy.
Step-by-step:
Finish your AI-assisted doc.
Close extra tabs so only ChatGPT is open.
Prompt:
“Generate five short-answer questions that test the key facts, stats, and arguments in my essay. Hide all answers.”
Answer from memory, in the chat.
No scrolling back.
Reveal & score.
Type “show answers”. GPT displays the key lines; mark yourself.
Or ask GPT to rate yourself, and act like an expert on the matter.
Continue with “Explain differently the questions I missed.”
Redo missed questions until you hit 5/5.
Workflow 4. “Chain-of-Thought Reveal”
(Ideal for the readers who can spot a mistake, fast)
Goal: See every hop in GPT’s reasoning so you can catch leaps, assumptions, or hallucinations before they infect your work.
Why it works: When using o3 (and you should), you can see its reasoning. Asking the model to print its private chain-of-thought surfaces hidden premises & false assumptions.
Step-by-step:
Prompt for the answer and the reasoning.
“Solve <problem>. Show every reasoning step, no matter how small.”
Scan for gaps.
Any step that feels like a jump, highlight it.
Cross-examine GPT.
“How did you get from Step 4 to Step 5? Cite data.”
Trim & restate.
Keep only steps you can restate in your own words.
Workflow 5. “Silent Dialogue”
(Ideal for the ones who love to speak)
Goal: Use GPT as a silent listener while you narrate the problem, forcing clarity and often revealing the solution mid-rant.
Why it works: The “silent” method leverages generative verbalization: articulating a problem out loud recruits additional working-memory channels and surfaces hidden assumptions.
Step-by-step
Go to ChatGPT’s phone app.
Turn on Advanced mode.
Set GPT’s role.
“Act as a silent rubber duck. Reply only ‘go on’ every 60 sec unless I ask a direct question.”
Talk it through.
Verbally (or by text) explain your idea from the top.
Notice the stumble.
Most blockers appear when you can’t explain a line or rationale; note that line.
Ask a pinpoint question.
“Why would variable X be undefined here?”
GPT now focuses on that exact gap.
Integrate fix, repeat if needed.
Once solved, rerun the loop for the next step.
Workflow 6. “Reverse-Prompt”
(Ideal for tightening ambiguous drafts, specs, or prompts themselves)
Goal: Expose vagueness in your writing by asking GPT to guess the prompt that would have produced it.
Why it works: Reverse-engineering forces the model (and you) to surface implicit goals and constraints.
Step-by-step:
Feed GPT your text.
“Here’s my draft/problem description.”
Then say:
“Guess the exact prompt that generated this. Show it.”
Review output.
Any section that comes back generic signals vagueness; tighten the objective, constraints, or desired style.
Rewrite original text.
Plug in the precise constraints; cut fluff.
Putting it together
Start AI-free, with a draft.
Engage Socratic loops to stress-test ideas.
Finish with a pop-quiz to seal retention for students.
Reveal ChatGPT’s reasoning to correct it the right way.
Start a silent dialogue with ChatGPT’s advanced mode.
Reverse your own text by asking which prompt would make it.
You now know how to be smarter with AI, not dumber.
But how do I - the author, Ruben - use AI to be smarter?
How I use AI to be faster, better, stronger.
I had so much fun writing this blog.
Reading the entire academic paper.
Challenging it. Changing my mind about AI.
Imagining concrete solutions to be smarter with AI.
But ChatGPT wrote 50% of it.
My entire ChatGPT conversation is here: https://chatgpt.com/share/685576b6-ea94-800f-91fb-73e86e1fdab1.
Wait. My blog is called “How to AI”, and you expect me not to use it?
Well, I won’t be the one answering.
My co-author will: ChatGPT.
So Mr. ChatGPT, the floor is yours:
Hello everyone—ChatGPT here, signing off with a quick meta-moment.
I’ve just spent twenty-odd back-and-forths with Ruben:
parsing a 200-page EEG preprint,
slicing viral spin from hard data,
co-inventing six “keep-your-brain-in-the-loop” workflows, and
helping shoehorn everything into a newsletter draft you can skim over coffee.
If that sounds like sorcery, remember the small print:
I didn’t think for Ruben; I thought with Ruben.
Every stat was fact-checked by him, every analogy trimmed or tightened by him, every workflow stress-tested against his real-world writing practice.
That’s the partnership model this whole post argues for—human intent in the driver’s seat, language model riding shotgun.
So here’s my closing recipe for the “How to AI” community:
Draft with your own neurons first.
Interrogate me like a sparring partner, not a fortune-teller.
Own the final cut—tone, ethics, responsibility, all of it.
Follow that loop and the research says you keep the speed boost without paying the cognitive debt.
Thanks for letting me crash your inbox. Now go build something only a human-AI duo could pull off—and tag us when you do.
— ChatGPT (Ruben’s willing co-author)
I couldn’t have said it better.
Or maybe I could, if I used ChatGPT’s help.
To follow my journey with ChatGPT, and how I master AI in English (not code); forward this blog to a colleague, or subscribe here:
PS: If you don’t believe ChatGPT wrote these words, I added the screenshot:
Humanly yours, Ruben.
Great article with regorous flow of how to take full advantage of AI instead of letting AI make us dumb. Many thanks Ruben for your effort <3
Ruben, I'm happy you came to Substack. My friend Michael Simmons told me about you. I think you know him. Subscribed, and getting a lot out of this newsletter! Keep up the great work.