What I Actually Lost When I Let ChatGPT Write My Code Comments

October 5, 2026

cognitive-debt ai-collaboration metacognition code-review llm-reasoning

Last month I couldn’t explain a function I’d shipped the week before. Not a complicated one — a rate limiter, maybe forty lines. I opened it to fix an edge case and stared at my own docstring like it belonged to a stranger. It described the backoff strategy, the reasoning for the jitter window, the tradeoff against a simpler fixed-delay approach. All correct. None of it mine. I’d asked the AI to write the comments, read them once, nodded, and shipped.

That nod is the whole problem. It felt like agreement. It was actually the moment I stopped thinking.

There’s a study out of MIT that’s been rattling around my head since I read it — Kosmyna’s “Your Brain on ChatGPT” work. Researchers had students write essays under three conditions: unaided, with a search engine, and with an LLM, while wearing EEG caps. The LLM group showed the weakest neural connectivity of the three, and a chunk of them couldn’t quote their own essay back sixty seconds after finishing it. Fifteen percent said the essay didn’t feel like theirs at all. What got me wasn’t the memory loss — it was the ownership number. That’s exactly the sensation I had staring at my own rate limiter.

The researchers call the buildup “cognitive debt,” and they’re specific that it’s not like financial debt — there’s no interest-only period where you’re fine. Kosmyna’s line was blunter than I expected from an academic paper: there’s no cognitive credit card, you can’t pay it off later. You either did the thinking or you didn’t, and the essay/function/document only tells you which one happened if you interrogate it before you accept it.

Here’s the reframe that’s actually useful, though, because “just think harder” isn’t advice, it’s a scold. The MIT researchers who study why LLMs fail at things they seem good at — Rambachan’s group, testing whether models build real world-models — found something concrete: an LLM can give you flawless driving directions through Manhattan and then crash spectacularly the moment you introduce one detour. It never had a map. It had excellent pattern completion over text that looked like directions. My comment block was the same kind of artifact. It looked like reasoning because it had the shape of reasoning — tradeoffs stated, alternatives dismissed, confidence expressed in the right register. Fluency is not the same thing as a claim that’s been checked, and I let the fluency stand in for the check.

Lev Tankelevitch at Microsoft Research has a frame I keep coming back to: using an LLM well is a management skill, not a prompting skill. Before you delegate anything to a report, you already know what you want, you’ve broken the problem into pieces small enough to hand off, and you know enough to tell a good answer from a plausible-sounding wrong one when it comes back. Skip that groundwork with a human report and you get sloppy work you rubber-stamp because you can’t tell the difference. Skip it with a model and you get the same thing, except the model never pushes back, never asks a clarifying question unless you make it, and writes with total confidence regardless of whether it actually understood the constraint you were dancing around.

The part I got backwards was sequencing. I was using the AI to arrive at the reasoning, then reading its output as a substitute for having done that arriving myself. The fix isn’t “write your own comments” as a purity ritual — it’s deciding, before I open the chat window, which parts of this problem I need to hold in my own head so that I can recognize when the model’s version of it is subtly off. The rate limiter’s jitter window was a decision with a real tradeoff behind it — throughput versus thundering-herd risk — and I’d let the model make that tradeoff, explain the tradeoff, and get credit in my own memory for having made it. Three separate acts, and I’d only done the last one, passively.

I don’t think the answer is fewer AI-written comments. It’s noticing that the moment you feel the flicker of “yeah, that’s right, good” while reading model output is the exact moment you should slow down and ask whether you’d have written the same sentence, for the same reason, cold. If you can’t, you didn’t approve the reasoning. You approved the fluency.

What’s still unclear to me is whether this gets easier or harder as the outputs get better. A worse model would have forced more scrutiny by being obviously wrong more often. A better one removes that friction entirely — and friction, it turns out, might have been doing more work than I gave it credit for.

Sources

We Taught Humans to Think Better Around AI. We Forgot to Teach AI to Think Better Around Itself.

April 16, 2026

ai-architecture metacognition uncertainty llm-design ai-reasoning