I used AI to learn Golang in 2025. I was faster than ever, but couldn’t remember what I’d written. A recent MIT study explains why.
Intro: An MIT study about our brains on ChatGPT #
A recent MIT study shows that using AI like ChatGPT can lead to “cognitive debt” . A kind of mental shortcutting that makes you faster now, but possibly weaker later.
I didn’t need EEG data to know this. I lived it.
In 2020, I learned Python without AI. In 2025, I learned Golang, and AI ended up doing 60% of the work.
Here’s what I’ve noticed.
Being faster and more efficient doesn’t seem like a tradeoff, but there are definitely some downsides to this.
TL;DR: What the Study Found #
Researchers asked people to write essays with or without ChatGPT, then measured their brain activity (EEG) and tested their memory.
Results? AI users wrote faster and felt more productive — but couldn’t recall what they wrote. Their brains were less engaged. Neural activity dropped.
The researchers called this “cognitive debt”: you get results fast, but you might be skipping the mental work that leads to learning.
It’s very interesting to me that the term cognitive debt was coined in this study to describe this. Just as technical debt (Or the creation of) produces fast results while creating “interest” by reducing future flexibility and increasing maintenance drag, cognitive debt.
My Coding Experience Mirrors This #
- Python in 2020: No AI, lots of struggle, but I still remember how most things worked. I own that knowledge. I’m using Python as an example, but I experienced this process with other scripting languages.
- Golang 2025: I started learning it on May this year. With AI, I ship faster. But if you asked me to refactor a function I co-wrote last week, I’d struggle a bit.
- The shift: I don’t really code much anymore. I describe what I want, the AI fills in the blanks, and I tweak the result. It works, but I can already feel the price.
This is exactly what the study describes: faster output, less brain engagement, and weaker memory encoding.
But it doesn’t stop with coding, obviously. The more I rely on AI for other things, the more lazy I become. I started developing a potentially toxic heuristic:
Ask AI first, then refine.
The New Layer of Abstraction #
This isn’t the first time we’ve offloaded complexity. In fact, this moment reminds me of how programming evolved over the decades:
- We wrote assembly/machine code in the 60s (Lower abstraction).
- We started with C/Pascal (Medium abstraction) in the 80s.
- We started with Python, Ruby, etc. in the 2000s (Higher abstraction).
- We’re starting now with prompting, a very high abstraction.
We’ve always abstracted away from hardware, toward ideas. We stopped writing bits and started writing logic. Now, we’re barely writing code — just intent. This is super interesting.
I haven’t written much code in the past few months. I just prompt, then review, tweak, ship.
Prompting feels like the next phase: writing pseudocode in English, and letting the machine turn it into implementation. I’ve read prompt engineering is the new technical literacy, and for a while I rolled my eyes at that statement. But now I see it as being potentially true.
I think this is a good thing — with a warning.
The Catch #
Is AI good for writing complex pieces of software and achieving auto-maintained, self-healing, applications? Definitely not (For now…). As with every AI evolution, I tend to think that what it can do poorly now, it’s just a matter of time for it to become good at it.
- GPT-2 wasn’t useful. GPT-3 changed everything.
- Early image models were memes, party tricks. A year later, people are making short films.
- AI can’t write production code yet. But give it time…
That’s why this shift feels like more than a trend. It’s a new abstraction layer. And like every leap before it, it asks less of our syntax memory and more of our creative thinking.
We’re moving from needing code to needing better ideas. And that’s one of my most important personal insights on all this. We need better questions. Better problems to solve. Better thinking. We can address this “gaps” or problems way faster than before and we’re beginning to bridge the gap between “how-to” and “what-to” or “why”.
The Middle Ground I’m Aiming For #
I now try to use AI more intentionally:
- I sketch the logic first, or pseudocode.
- I try to implement tricky parts and build the code skeleton myself.
- Then I use AI to clean up, debug, or optimize. Using it like an idea autocompletion tool.
That way, I stay engaged but don’t waste time reinventing boilerplate.
The same process goes for non-code (Writing documentation, research, learning new topics, etc.).
Where This Leaves Us #
AI is a tool. And like all tools, it shapes us, the people who use it.
Used with intention, it can help us think faster, solve faster, build faster.
But used too early, or too often, it risks turning us into managers of output we never deeply understood. We might get results, but miss out on learning.
You could wake up one day with dozens of finished projects… and no clue how they work (This makes me scared a little bit, to be honest).
Conclusion: Your Brain, Your Call #
Tools like ChatGPT, Claude Code, Codex, Windsurf, etc. are amazing — but they’re just that: tools.
Use them with awareness. Especially if you’re learning.
Final Thoughts #
This is what the MIT study showed, and what I’ve lived firsthand:
- AI makes us more productive in the short term.
- But it can make us less engaged, less aware, and less connected to the work we’re doing.
- There’s a sweet spot: start with your own brain. Then refine with AI.
If you’re using AI to learn, build the mental muscle first. Let the machine lift after you’ve done the heavy thinking.
Prompting is just the next phase in abstraction but clear thought matters. A lot.
