I write about AI a lot. Sometimes for the companies building it, sometimes for people studying it, sometimes for the general public. I’m quite optimistic about technology. But now, nearly two years after ChatGPT’s release, I’m starting to feel something I didn’t expect: fatigue. Not just with the tool itself, but with having to defend it.
It’s just so… everywhere. In places where it doesn’t really have to be. It’s like 2016 Instagram filters all over again. The question isn’t “Did you use it?” It’s “Why did you need to?”
By the time ChatGPT launched, I’d already been writing professionally for a few years. It wasn’t my entire job, but it was a big chunk of it. So when I started using LLMs at first, it was fabulous. Writing articles that compare one SaaS tool to another used to take many hours, and now it took fewer hours.
I think like a lot of people, I felt something between awe and dread.
Fast-forward to now: AI tools are embedded into most workflows. Students use them for essays, professionals use them for emails, and LinkedIn influencers use them to sound exactly like other LinkedIn influencers.
So what comes next?
MIT Labs just published a 206-page study. Granted, the last 20 or so pages were just tables and references, so I skipped those, but I did read the rest.
A group of researchers led by Nataliya Kosmyna at MIT’s Human-AI Interaction Group set out to answer what might be one of the most important questions we have about LLMs today: What happens to your brain when you use ChatGPT to help you think?
Here’s the link to the whole paper for whoever’s interested.
Let’s beginnnnn.
Whose brains were being looked at?
Not your regular-degular folks. The study included 54 participants between the ages of 18 and 39, they were a mix of undergrads, grad students, researchers, business professionals, and engineers from MIT, Harvard, Wellesley, Tufts, and Northeastern (big fancy colleges).
These were people who knew how to write, had real jobs, or were pursuing degrees that required serious thinking. Some even identified as full-time writers.
Okay next,
Each participant completed four writing sessions over the course of a few weeks. They had to write essays on SAT-style prompts (so reasonably complex topics). What’s also cool is that it didn’t just pit AI vs. humans in a gladiator-style smackdown.
Instead, they split these people up into three different groups:
- Group 1: Write essays using ChatGPT. Use the LLM however you like. Ask for an outline, have it write the whole thing, get a paragraph and edit it, whatever.
- Group 2: Write essays using Google or Bing Search. Do your own research, but write the essay yourself. Also WHO uses Bing?
- Group 3: Brain-only. No tabs, no tools. Just you and the blank page. Good luck.
Side note: I found the study’s design unintentionally generational.
- ChatGPT is the tech I didn’t grow up with. It wasn’t part of school, college, or even the early years of working life. It’s new, and it changes the task itself.
- Search engines are the tools I DID grow up with. Writing essays meant jumping between a blank doc and a bunch of open tabs.
- Brain-only is how our parents probably did their writing. No internet, just paper and maybe a stack of reference books.
While these people were writing, they were hooked up with 32-channel EEG caps that recorded brainwave activity in real time. It would track things like attention, processing, and working memory.
Their essays were scored both by human graders and AI models. Later, the participants were asked how much they remembered, how confident they felt, and whether the essay felt like it was really theirs.
OH AND!!!!
They got only 20 minutes per essay. (TWENTY MINS ARE U KIDDING?) I have just spent twenty minutes deciding whether to put ‘?!?!!’ or just a single ? after the word kidding, but go off.
Finally,
The study used what’s called a crossover design. That means every participant went through all three writing conditions: once with ChatGPT, once with a search engine, and once with no tools at all. This matters, because if they’d just split people into three fixed groups and compared them,
You’d always wonder: Was the difference in results because of the tools, or because of the people? What if one group accidentally got all the good writers? What if one group had the slow typers? What if ALL the participants were pissed that the electrodes were ruining their one good hair day that week? (can I get an amen)
A crossover design solves that. By rotating each person through all the conditions, the researchers could isolate the impact of the tool itself, and not the quirks of the person using it. Every participant becomes their own before-and-after.
I’m gonna throw in a sports metaphor here just for practice:
If you’re a basketball coach trying to figure out whether Nike or Adidas improves your players’ jump shots, here’s what you don’t do: make one player wear Nike and the other wear Adidas, then compare their scores. No. Bad coach. That tells you more about the players than the shoes.
What you do is have each player shoot 50 times in Nikes and 50 times in Adidas. Same person, different shoes. Then you compare. That way, you’re testing the shoes, not the shooter.
That’s crossover design.
All good so far? Is everyone still here? Susu break? No? Okay let’s go on.
Let’s talk about EEGs for a second.
Now, if you’re not a neuroscientist (hi, same), you might assume EEGs are just the default, the gold standard. They look pretty legit.
But that’s not exactly the case. They’re more like the duct tape of brain research. They’re reliable and not too expensive.
EEGs are great at capturing WHEN your brain does something. They can pick up shifts in attention, memory strain, and cognitive effort down to the millisecond. That’s their strength: timing.
What they can’t do well is tell you very precisely WHERE in the brain that activity is coming from. The signal has to travel through your whole head. That includes hair, silicone build-up from too much conditioner, your scalp, multiple layers of tissue, bone, some watery stuff, and then layers of your actual jelly-brain. So the location data is actually not super precise.
Now, technically, you could use something fancier. Like fMRI, which gives you beautiful, high-resolution images of your brain lighting up. But that requires lying flat inside a giant magnet, perfectly still, with your head bolted in place. You can’t write an essay in those conditions.
EEGs, on the other hand, let you sit at a desk, open your laptop, and write like a normal person with weird headgear.
The reason I took this little detour is because study design matters. When you read findings like “ChatGPT users had lower memory retention,” it’s worth asking how that was measured, and what kind of tools were used to get that data. It’s not always the best tools, it’s always the most appropriate given what’s available. Just worth keeping that in mind.
Special thank you to my neuroscientist friend Aditi Karmarkar for reviewing this whole section and pointing out some other helpful stuff to make this piece better.
Now, let’s get to the findings
Let’s start with the EEG data. They tracked four kinds of brainwave activity. Certain frequencies are strongly associated with different kinds of mental effort:
- Alpha waves suggest your brain is retrieving and connecting ideas (semantic processing)
- Beta waves show focused attention (what you use to organize an argument or plan your next sentence)
- Theta is associated with working memory (keeping several ideas in your head at once)
- Delta reflects deeper, large-scale cognitive integration (when your brain is trying to make something make sense)
When participants wrote their essays using ChatGPT, all four of these brain waves dropped. Alpha and Beta especially. That means less idea generation, less mental focus, less work, period. That doesn’t mean people were being lazy. But it does mean they were thinking less because the tool was doing the heavy lifting, cognitively speaking.
The brain-only group showed the opposite: high activation across the board. Writing without support forced participants to retrieve knowledge, form arguments, and phrase everything from scratch. Unsurprisingly, that kind of effort requires a lot of brain. It was the most cognitively demanding condition.
The Google search people landed in the middle. They still had to write their own arguments, but they had the comfort of looking things up, checking facts, or drawing from others’ phrasing. The burden wasn’t quite as heavy as doing it all by themselves.
Now if you had to guess: Which group was happiest with their essay?
Was it the brain-only group, who did everything on their own and walked away with that warm glow of authorship?
Or the ChatGPT group, who got smooth, polished writing with minimal friction. What’s not to like?
Or was it the search group, who had access to resources but still had to write the thing from scratch?
Here’s how it actually played out:
- Search group: EVERYONE said they were satisfied
- ChatGPT group: 17 out of 18 were satisfied
- Brain-only group: 15 out of 18 were satisfied. Three were unsure or disappointed.
Huh.
It’s tempting to dig for a big lesson here, but what I think is that people like what they know.
For this specific age group (late teens to late 30s) search engines are the default tool for writing. That’s how they grew up doing assignments. It’s a cozy workflow. Look something up, open a hundred tabs, frantically put stuff down on a document, borrow a stat, mash together an argument, slap a conclusion. Done.
ChatGPT, on the other hand, is newer. More powerful, but also less predictable. To get something genuinely satisfying out of it in 20 minutes, you need to prompt well, edit fast, and be really clear about what you want. Prompt engineering is a skill that not everyone has figured out yet.
The brain-only group had no lifelines. For people under 40, that mode of writing is borderline alien. It’s how your parents wrote essays (and I don’t know how much our parents loved it either.)
Also worth mentioning, with a sample size THIS small, I don’t think these results are statistically significant so it’s probably not worth drawing any major conclusions from this.
So what happened in each group?
Let’s start with memory.
Minutes after every writing session, participants were asked to recall a sentence from their own essay and write it down. Simple enough. You wrote it. Can you remember it?
In the brain-only group, everyone could. All 18 nailed it.
In the search group, two people couldn’t remember their quote accurately.
In the ChatGPT group, four people couldn’t recall what they’d written. At all.
Four people may not seem like much, but that’s not nothing. Especially because the essays had just been written minutes earlier. It wasn’t a long-term memory test. The only difference was how involved your brain was in producing the sentence in the first place. The less effort you put in, the less likely you were to remember the result.
This tracks with what the EEG already showed. When the cognitive load drops, so does retention. And in the case of ChatGPT, a lot of the cognitive work was taken care of.
I’d say in most conditions when you’re given 20 minutes to write an essay, you’re more of a sentence curator than a sentence builder. Anyway, so it makes sense that when it came time to remember the sentences you wrote later, there was no construction memory to pull from. You can’t recall the steps of a process you didn’t go through.
If you’ve ever used ChatGPT to write something, then forgotten it five minutes later, this is why.
Next, did this feel like yours?
The study also looked at perceived authorship.
In the brain-only group, the answer was a pretty confident yes. Sixteen out of eighteen said they felt full ownership. The other two were still mostly on board, just noting that their ideas were probably shaped by past reading or general exposure. Fair.
In the search group, it was mixed. Only about a third said “this is fully my work.” The rest gave more of a shared-credit response. “Some of this is me, some of this is stuff I found online.”
And the ChatGPT group of course was all over the place. Some said “Yep, this is mine.” Others said “Not at all.” And then there were the delightful math answers: “70/30 between Chat and I,” “maybe 60% AI, 40% me,” etc. One participant described it as feeling like “being the editor of someone else’s voice,” which is… pretty accurate.
Now, this doesn’t mean the ChatGPT group just sat back and did nothing. Some participants rejected what the model gave them and rewrote large chunks. Others kept re-prompting until the tone felt right. One person described feeling “stuck” because there were too many good options and they didn’t know which one to go with.
The work shifted from less writing to more managing the writing.
Also worth flagging: Even when people did prompt well, the study noted a drop in delta waves (those are the ones associated with deep integration of ideas.)
Meaning even when ChatGPT is helping you produce clean, organized text, it doesn’t mean your brain is putting the same kind of work into connecting the dots or understanding the relationships between concepts.
FINAL STUDY TAKEAWAYS:
Across all these metrics (memory, attention, semantic integration) ChatGPT use consistently lines up with lower brain activity. Even when people were editing, or second-guessing the output, they weren’t activating the same regions.
Although the essays came out structured, readable, and polished, the participants using ChatGPT remembered less, felt more disconnected, and showed less cognitive engagement.
Oh, and in case you’re wondering. Yes, the ChatGPT essays scored the highest.
When human graders and AI models scored the essays, the ones written with ChatGPT generally came out on top. They were more polished, better structured, and used more complex language. They especially stood out on things like logic, cohesion, sentence variety, and vocabulary.
Which makes sense. ChatGPT is trained on billions of examples of good writing. It knows how to sound smart. If you’re measuring how clean and professional something looks, especially under 20 minutes, it’s hard to beat.
But scoring well doesn’t mean you actually did the thinking. Just like using a calculator doesn’t mean you understand how to solve the problem. And how much that matters to you depends on what you think writing is for. If it’s just about getting a task done well, sure, let the tool help.
But if your goal is to clarify your thinking, develop your ideas, or become a better writer, there’s a cost to outsourcing the hard part.
Let’s take a break here
In 1790, Immanuel Kant warned that enlightenment was not the possession of knowledge, but the courage to think for oneself.
You don’t become a better writer by producing words. You become a better writer by making decisions. What do I actually want to say? Why does that matter? Is that even true? Is there a better way to say it? Does this argument fall apart?
When your invisible thoughts become visible, you are forced to wrestle with them in reality and not your imagination. You can’t wave your hands and say “you know what I mean” because you have to spell it out. And if you can’t, it becomes obvious that maybe YOU didn’t know what you meant.
That process is frustrating, and very often, it’s super humiliating. But when you stick with it, the rewards are MASSIVE. You get so, so smart. You learn to spot lazy reasoning. You develop a feel for structure. And you stop being fooled by smart-sounding sentences that don’t say anything.
AI doesn’t stop you from doing that. But it does make it easy not to.
ChatGPT gives you fast, fluent answers that are reasonable but not really surprising. It won’t flag your bad reasoning. It won’t force you to question an assumption. And if the output looks good enough, you’ll probably just go with it.
Because the model is trained to be average (literally. statistical likelihood is its job), it won’t offer anything strange, risky, or new. You won’t learn from being wrong. You’ll just get faster at sounding okay. But if you’re trying to become a better writer or a smarter thinker, the process IS the point.
If you skip it too often, you lose the muscle that tells you when something’s worth saying.
Stop, and think of something where the process matters, not just the outcome
Quality is never an accident. It is always the result of intelligent effort.
– John Ruskin
There’s a lot I want to say about AI and writing, but I’m not going to cram it all into one essay. Mostly because I don’t want to. Also because I probably shouldn’t.
We’re living through a weird moment where a very powerful tool has landed in our laps, and most of us are too busy using it to actually think about what it’s doing to us. So instead of hot takes and sexy conclusions, I want to slow down and think about this a little more intentionally.
With that in mind, I’ve been thinking about something pretty simple: when does the process actually matter more than the result? Or matter as much? Or matter enough not to skip it.
What if you could instantly download any skill into your brain. No trial, no error, no learning curve. You just… have it.
In some cases, it would be amazing. Like I would love to download a new language into my brain. The process of learning a new language is dumb and I can’t think of any benefits to sounding like a buffoon while you fumble through bad pronunciation.
Another example is if you have to write a long, detailed report on something. Barely anyone reads those, and unless you have to present it to people who may ask you questions, I don’t think there’s some glorious benefit to writing it all by yourself.
But I started thinking about the stuff where the process actually matters. Situations where if you skip it, you don’t build the instincts that make the skill usable in real life.
Here’s a working list:
- Riding a cycle. If you never fell, you’d never learn how to react when things go wrong, which happens a lot with cycling.
- When you’re learning to cook and you screw up, you learn a lot of interesting hacks to fix things. For example if something is too spicy you can add some sour to cut the edge off. You learn to salvage stuff.
- Communicating with people. If you don’t make mistakes, you skip some really important learning, like how to apologize. How your partner prefers being apologized to. How they react when they’re annoyed with you. How they like to be appreciated. Some of that stuff only comes up when there’s friction and imperfection.
- Parenting. You can know everything about sleep schedules and child development and still be bad at parenting. The process teaches you how to CARE and LOVE under pressure, which is really all it comes down to.
- Developing convictions and beliefs. Unless you bump up against all your inconsistencies and wrestle with your doubts, you aren’t ever truly firm in your beliefs. If you just hit the finish line, you’ve skipped the part where you strengthen the foundations of your convictions.
- A sense of humor. I cannot overemphasize how important this one is. You need to learn how to ruin a joke, how to recover, and how to read the room.
- I think getting over a heartbreak? Unsure. A shortcut would be great, but the pain teaches you how to be alone, how to grow, how not to do it again (hopefully).
This is obviously a vague and messy list. I’d love to hear what else you’d add to it based on your experience.
Convenience, meet consequence
At this point, it’s customary to offer a vision of the future where balance is restored. Where we use AI responsibly. Where schools teach “critical AI literacy.” Writers still write. Thinkers still think. And the machines politely assist.
That’s not where this is going.
Let me be clear: I’m not a luddite. I like AI. I’m fascinated by it. I will absolutely let it summarize a meeting transcript while I shop online. I’m not trying to go back to a typewriter and a bottle of whiskey.
But I’m also not excited about a future where we’re incentivized to outsource the one thing that actually matters: the part where your brain does its best work.
Most people, given the choice between “hard and meaningful” vs. “easy and good enough,” will pick the second one every time. I would too. So the question isn’t “Will AI ruin learning?” It’s “Will we still choose to think when we don’t have to?”
That’s not a rhetorical question. And I don’t know the answer to it. I just think that if creativity turns into prompt-engineering, the true scarce resource will become our beautiful, deranged, and unprompted imagination.
Unfinished Conversations 
Leave a comment