Skip to content
amond.dev
Go back

The Paradox of LLM Advancement

Updated:

The easier writing becomes, the more we forget what we wrote


It’s been more than two years since ChatGPT came out. Writing has gotten a lot easier in that time. I hand off work emails and document drafts to AI more often than I used to.

But then a thought hits me. That report you had Claude write yesterday — do you remember what it said?

As writing gets easier, are we actually losing the ability to understand what we write — even to remember our own words?


A concept called cognitive debt

Nataliya Kosmyna’s team at the MIT Media Lab published a study in 2025. Titled “Your Brain on ChatGPT,” it tracked 54 people across three groups over four months of essay writing.

The results are a bit bitter. 83% of LLM users couldn’t properly quote from the essay they had just written. The group that wrote without tools remembered theirs almost perfectly.

The researchers called this “cognitive debt.” It’s a close cousin of the “technical debt” developers talk about: piling up long-term cognitive cost for short-term convenience.

“The LLM group felt little ownership of their work and couldn’t recall the content even right after writing it.” — MIT Media Lab research team

The EEG measurements told the same story. ChatGPT users showed lower activation in brain regions responsible for memory encoding, attention, and executive function. In the researchers’ words, the brain was “literally switched off.”

What worries me more is that it becomes a habit. When people accustomed to LLMs were asked to write without tools, they couldn’t focus as well as those who had written unaided from the start.


The paradox of digital natives

We call Gen Z and Gen Alpha “digital natives.” They grew up with smartphones from birth, after all. Yet recent studies show rather unexpected results.

According to the EU Commission’s International Computer and Information Literacy Study (ICILS), digital literacy among 14-year-olds has actually declined since 2018. 43% of students possess only limited digital competence.

Research from the University of Stavanger points to a more fundamental problem. About 40% of Gen Z is losing the ability to communicate in handwriting. University professors say students avoid complex sentences and write only short ones, like social media posts.

GenerationTraitParadox
Gen ZDigital devices from birthWeak grasp of basics like local files and folder structures
Gen AlphaGrowing up with AIConcerns over declining reading and writing skills

Handling digital devices fluently and understanding and producing writing turn out to be separate skills.


The more you know about AI, the less you verify

One study presented at CHI stuck with me. It found that users with higher AI literacy — people who understand well how LLMs work — actually fact-check less.

Isn’t that ironic? The confidence of “I know how AI works, so I can use it well” is exactly what makes people skip verification.


So what should we do

The MIT researchers offered one clue. The group that first wrote without tools and brought in the LLM later showed much higher brain activation. They used AI as a tool to augment their ideas, not to replace them.

In other words, timing matters.

PrinciplePractice
Think firstOrganize your own thoughts before writing the prompt
Review criticallyDon’t use AI output as-is; revise it
Manage systematicallyOrganize generated content and revisit it regularly
Practice independentlyKeep writing without AI

I’m trying to follow this myself these days. I sort out my thinking first, then let the AI handle the polishing. If the order flips, it stops being my writing and becomes the AI’s.


Closing

It’s true that LLMs have made writing easier. But it also feels like we’re losing something in proportion to that ease.

In an age when writing is easy, what we may need is the skill of remembering what we write. Or maybe just writing without AI once in a while.

I wrote this post with Claude’s help too. But I drafted the outline first, and I decided the structure. Maybe that’s why I remember what it says. Probably.


References


Share this post on:

Next Post
A New Bar for AI Coding Benchmarks: The Question FrontierCode Asks