AI and Cognitive Debt: What I've Learned Using It Daily

Whether we like it or not, AI has burst into our work environment and disrupted the status quo we'd maintained for years. I currently work as a Staff Engineer with seven teams under my umbrella, and I use AI every day: cross-cutting analyses, spikes where I try different approaches before committing to one, architectural decisions that impact multiple teams, and a long list of other things. And something I've realized is that the machine already comes up with more options than I do for certain problems. Options that hadn't even occurred to me.

This has caused a thought to gradually take hold in my mind. Am I becoming a better engineer, or am I becoming dependent?

It's very tempting to let the machine do everything. To let it decide what to analyze, how to analyze it, and what conclusions to draw. And when you've been using it intensively for weeks, the line between "AI helps me think" and "AI thinks for me" becomes very thin.

When I discuss this with colleagues, I see they share similar concerns. A growing sense of becoming dependent. And a fear I hadn't heard before: that if tomorrow for whatever reason the token price skyrockets or the tool becomes unavailable, we won't be able to get the work done on our own anymore.

What Happens to Your Brain When You Use an LLM?

Researchers at the MIT Media Lab decided not to stop at surveys or subjective opinions. They ran a four-month study measuring participants' brain activity with electroencephalograms.

The study ("Your Brain on ChatGPT", Kosmyna et al., 2025) took 54 participants and divided them into three groups. Each group was assigned a different tool: one used ChatGPT, another a search engine, and another only their brain. The task was writing essays, across three sessions. In the fourth they reversed roles, making the ChatGPT group write without any tool, and the brain-only group switched to ChatGPT.

These were the results:

• Brain connectivity systematically decreased the more external support the group had. The brain-only group had the strongest neural networks, in contrast with the ChatGPT group.
• ChatGPT users couldn't cite the essays they had just written. They didn't remember what they had "produced."
• When ChatGPT users tried to write without AI in the fourth session, they showed weaker neural connectivity than those who had never used it. That is, the effect doesn't disappear when you put the tool down.

Additionally, this study was conducted before the arrival last November of Claude Opus 4.5 or Gemini 3.0, which marked a significant leap from previous models.

The researchers call this cognitive debt. AI saves you mental effort now, but you end up paying the consequences in the long run. Reduced critical thinking, diminished creativity, superficial information processing.

I'm sure this sounds familiar. You generate code with AI, commit it, and the next day you don't remember what that code does or why you wrote it that way.

OK, but That's with Essays. What About Learning?

Hamsa Bastani and her team at Wharton/UPenn asked themselves the same question. They designed a controlled trial with nearly 1,000 math students (published in PNAS, 2025). Three groups:

1. GPT Base: free access to GPT-4, with no restrictions. Students could ask whatever they wanted and copy answers directly.
2. GPT Tutor: access to GPT-4 but with guardrails designed by the professors. Instead of giving direct answers, the AI asked questions back, gave hints, and guided students to reach the solution on their own. This is known as the Socratic approach, after Socrates' method of teaching through questions rather than statements.
3. Control: no AI. Only textbooks and notes.

During the practice phase (with AI), the results were very good. GPT Base performed 48% better than the control. GPT Tutor, 127% better. Up to this point, everything seemed to indicate AI helped.

Then came the exam. Without AI.

The GPT Base group performed 17% worse than those who had never had AI. Not 17% worse than with AI, but worse than if they had never used it. And not only that — those same students were excessively optimistic about their abilities. They believed they had learned, but when the tool was taken away, it became clear they hadn't. This is what the literature calls the illusion of competence. AI's results mask your real gaps, and you don't realize it.

In contrast, the Socratic tutor group did not have this negative effect. They performed better with AI and didn't decline without it. There lies the crux of the matter: the effect doesn't depend on whether you use AI or not, but on how you use it.

And What About Developers?

The two previous studies involve students and essays or math problems. The obvious question is whether this applies to software engineers too.

In January 2026, Anthropic published a controlled trial with 52 software engineers, mostly juniors. They split them into two groups, one with AI access and one without. The task was to learn a new Python library (Trio, for asynchronous programming), complete a series of exercises, and then take a quiz to measure how much they had actually understood.

The AI group scored 17% lower on the quiz (50% vs 67%). The difference was especially large on debugging questions, which measured the ability to understand when code is wrong and why. In other words, those who used AI to learn were worse at precisely what you need most when something breaks in production.

But the most interesting part of the study isn't the overall result — it's that Anthropic analyzed how each participant used AI. Not everyone who used AI scored poorly. It depended heavily on how they interacted with it.

Low-performance patterns (average <40%):

• Total delegation: they asked AI to write all the code. They finished exercises faster, but then couldn't answer anything on the quiz. They hadn't learned a thing.
• Progressive dependency: they started trying to do things on their own, but gradually delegated more and more until AI was doing everything.
• Iterative debugging with AI: when something failed, they pasted the error into AI and applied the solution without stopping to understand what had happened.

High-performance patterns (average >65%):

• Conceptual investigation: they used AI only for conceptual questions ("Why does Trio handle concurrency this way?", "What's the difference between this and asyncio?"). They made many errors in the exercises but solved them on their own. This was the fastest pattern among high performers.
• Generation + comprehension: they let AI generate code, but then asked follow-up questions to understand why it worked and what would happen if they changed something.

One of Anthropic's conclusions in the paper is that "cognitive effort — and even getting painfully stuck — is likely important for fostering mastery."

So What Do We Do?

When I discuss these studies with colleagues, I usually see three reactions.

The first is "don't use AI." I don't think that's realistic or desirable. AI multiplies your capacity for analysis, research, and exploring options. Giving that up makes no sense.

The second is "use it for everything." But this is exactly what the three studies show doesn't work. If you delegate thinking, your brain works less (MIT), you learn less (Bastani), and you debug worse (Anthropic). And on top of that, you don't realize it's happening, because the illusion of competence makes you believe you're doing fine while you atrophy.

The third is "I'll get used to it." The data doesn't point in that direction. In the MIT study, after four months, those who used LLMs and tried writing without AI showed weaker neural connectivity than those who had never used it. The effect accumulates; it doesn't improve with time.

When AI Makes You Think More

Recently a couple of teams at the company dove headfirst into introducing AI into their products. One of the big challenges was how to validate that agents actually did what was expected of them — that they didn't hallucinate. They asked me for help.

I started by using AI to learn about the topic. I did research on how the industry is approaching agent evaluation, what approaches exist, what's mature and what isn't. With that knowledge in hand, I used AI to iterate on the different possibilities, libraries, implementations, pros and cons of each.

With all those options in front of me, I didn't just go with the machine's recommendation. I implemented several. I tested them. I saw firsthand how each one looked, what implications it had. And I iterated, shaping the implementations, deciding myself what structure made sense so the other teams could use and maintain them. The machine reached its conclusions, but I was the one who had to validate or correct them and steer the solution toward where I wanted to go.

In the end we had several approaches and reference implementations working in a test environment, which we brought to the teams so they could analyze which fit best in their context. In the past, something like this would have taken me much longer, or I simply wouldn't have done such a deep analysis with so many possibilities on the table.

I wasn't faster, but I made better-informed decisions.

When AI Makes You Think Less

But it's not all roses.

A while back I was implementing a user story on a product and gave the machine too much freedom with the implementation. This resulted in it creating a test suite that wasn't actually validating the real code implementation. The implementation could have been wrong, and those tests would have passed green anyway. It wasn't until a colleague called me out on it that I realized.

That could have caused a production incident. For not paying the attention it deserved. The AI had generated something that looked right, and I hadn't stopped to check whether it actually was.

What I've Learned

If I look at the three studies and what's happened to me, it all points in the same direction. AI amplifies what you're already doing. If you think, it helps you think more. If you delegate, it helps you delegate more.

With all of this on the table, I try to follow a set of rules to avoid falling into this trap:

Mechanical tasks: for AI. Formatting, generating boilerplate, routine refactors. There's no cognitive loss there.

Amplifying analysis: I ask it to give me more options, more depth, angles I hadn't considered. But I evaluate the options myself and don't accept the first thing it gives me. In the agent evaluation spike, I implemented several options precisely to compare firsthand, not to go with the machine's recommendation.

Making decisions myself: that I don't delegate. Recently two teams came to me to resolve an integration that affected both of them. First I thought about it in my own head, asked the involved parties for more context, and mentally worked through several options. Then I dumped all that knowledge into AI: the context, the problem, why it was a problem, and the options I had thought of with their pros and cons. I asked it to look for more information, to challenge me, to propose additional options using mine as a starting point. The AI made a recommendation, but the one with the full context was me. I was the one making the decision and taking responsibility for it.

I keep seeing people on LinkedIn who spin up who knows how many sessions or agents in parallel to crank out a ton of work at once. Honestly, I don't know how they do it. I'm unable to maintain that many contexts in my head at the same time, and forcing it means I don't give them the attention they need. For me, AI has become a tool for deep research, for going back and forth on a specific topic, for going much deeper than I did before — not for going faster.

The code that goes to production is your responsibility. It doesn't matter if AI generated it. If that code fails at three in the morning, you can't say "AI did it." You are the owner and the consequences are yours. And to truly own something, you have to understand it.

Ian Cooper, in his essay "Coding is Dead, Long Live Programming", talks about what he calls the "theory of the program." It's the deep understanding of why the code is the way it is. And he proposes a very simple test to know if you have that theory. Ask yourself if you can explain it to someone else. If you can explain it in your own words, it means you managed to retain the necessary context and understand what was done. If you can't, maybe you're delegating too much to AI.

To Wrap Up

A tool that thinks for you doesn't make you smarter. It makes you more comfortable.

I use AI every day and I don't think I'll stop using it in the short to medium term. But I've resolved that if tomorrow it's taken away from me, I still have to know. Still be able to debug my own code, make design decisions, explain to someone why the system is built the way it is.

It's a good question to ask yourself from time to time.