Skip to main content

TECH VEDA

Corporate on-site training - Submit proposal Pick your modulesSharpen your kernel skills: deep dives, drivers, Yocto, CVEs, careers — updated daily. Read the blog →
Career

Is AI Weakening My Critical Thinking as an Engineer?

AI can speed up routine work while the thinking you need for hard problems weakens. What developer studies show, what is unknown, and how to check yourself.

Is AI Weakening My Critical Thinking as an Engineer?

AI is making routine engineering work faster. The quieter questions are what it does to the thinking you need when the routine answer fails, and why that thinking gets harder to start. Here is what is known, what is not, and how to watch it in yourself.

Studies of developers show three short-term effects of AI help: less understanding forms, more wrong answers get through, and people feel more capable than they are. Research on effort and offloading suggests why it is hard to reverse: each hand-over makes the next one come sooner. Long-term skill loss in experienced engineers has not been measured, but two simple gauges let you watch both changes in yourself.

The code compiles, but you cannot say which lock is held

Last month an AI assistant helped you write a driver change. It works, it passed review and you can read every line. Today a colleague asks a simple question about it: in this function, which lock is held, and what happens if the interrupt fires here? You recognise the code, but you cannot answer.

That is critical thinking in engineering: small acts such as predicting what a register should read, asking why a fix works and doubting an answer that looks right. Our post on building a mental model describes the understanding they rest on. An assistant can do each of them for you, and each time it does, you skip a repetition.

What has already been measured in software engineers

Understanding forms less. In a recent experiment, not yet peer reviewed, developers learned a new programming library with or without an AI assistant. On a quiz straight afterwards, those who had used AI scored about 50% against about 67%. The largest gap was on debugging questions. Those who used it to ask for explanations, not only for code, did about as well as those without it. Our post on AI-assisted learning or avoidance covers that experiment.

Wrong answers get through. In an experiment with programmers, also not yet peer reviewed, they caught incorrect AI-written checks only about half the time, with similar confidence either way. In a security study, AI users wrote less secure code yet rated it more secure.

People misjudge their own work. In a 2025 trial, experienced open-source developers believed AI had made them faster when they were measured as slower. Newer data suggests the tools may now help, but the gap between felt and measured performance is the lasting finding. In a survey of knowledge workers, more trust in AI went with less reported critical thinking.

The pattern is consistent: understanding forms less, wrong answers get through, and people feel more capable than they are.

What has not been measured yet

No study has followed experienced engineers for years to see whether skill they already had erodes with heavy AI use.

Research on skills that fade through disuse finds that thinking-heavy skills tend to fade faster than physical ones. That is not a study of engineers, so treat it as a reason to watch, not as proof.

Why the mind stops wanting to engage

A second change explains why the first is hard to reverse. Working a problem through yourself costs effort, and people reliably choose a less demanding route when one is on offer. An assistant is always on offer.

Each hand-over also makes the next more likely. In one experiment, people who had used a search engine for hard questions later reached for it more often and more quickly, even for easy ones. In another, offloading made people faster at a task but left them remembering less of it. A small brain-recording study, not yet peer reviewed, found weaker engagement in people who had written with an AI assistant for several sessions and then had to write without it.

Together these suggest a loop, though this is our reading, not a measured result: you reach for the assistant sooner, so you practise less, so doing it yourself feels harder, so you reach sooner still.

The reluctance to think it through yourself is not a flaw in your character. It is a habit, and a habit can be measured.

What to expect over time: productivity and challenge capacity

The table is our own projection from this evidence, not a measured result, for an engineer who lets the assistant do most of the reasoning.

PeriodRoutine workNew or hard problemsWhat you might notice
First monthsFaster, more outputLittle change; existing skill carries youNothing, or feeling more capable
First year or twoStill fastSlower to form a hypothesis unaided; more trial and errorReviews feel easier; explaining your own changes feels harder
Longer termFast while the tool knows the territoryWeaker where the tool has no precedent: a new SoC, a rare raceHard bugs get handed upward or stall

Tickets closed go up. Your ability to reason through a problem nobody has solved is not tracked at all.

Routine productivity can rise while the capacity to handle hard problems falls, and nothing tells you it is happening.

That matters for your career: the hard problems, such as an intermittent boot failure, reach senior engineers, and an assistant is least likely to have seen them. Our post on epistemic debt describes a team that ships code nobody can still reason about.

A way to watch it: the reach log and the monthly baseline

This method is our own suggestion, not a tested one. It watches the loop from both ends: how soon you hand a problem over, and what you can still do alone.

1. The reach log, every working day. Pick two core skills, such as reading a kernel oops or finding the commit behind a regression. Each time one comes up, note when you started, when you first asked the assistant, and whether you had written down a hypothesis before asking. Writing the hypothesis first is the habit from our post on why work years are not practice years. When you feel the pull to ask, note what set it off: stuck after real effort, or the first moment of discomfort.

2. The monthly baseline, one hour a month. Take one closed problem per skill from your own history, of similar difficulty each month, and work it without the assistant for up to 45 minutes. Record whether you solved it and how hard it felt from 1 to 5, then check against the known fix.

MonthMedian minutes before askingAsks with a hypothesis firstBaseline solved?Felt effort (1–5)
1(fill in)(fill in)(fill in)(fill in)
2(fill in)(fill in)(fill in)(fill in)

3. Read the two gauges together after three months:

  • Both steady: carry on.
  • Minutes before asking falling, baseline steady: the habit is forming before the skill has gone, the cheapest moment to act.
  • Felt effort rising, baseline holding: the skill is still there, but engaging costs more. That is the resistance, not yet depletion.
  • Baseline falling too: depletion is under way.

In the last three cases, add one problem a week on that skill done unaided and checked with the assistant afterwards, the idea behind keeping one core skill in your own hands, and set a floor: no asking before a written hypothesis. These are rough proxies, not validated tests, so read the trend, not one month.

Watch how soon you reach for the tool. It tends to change before the skill does.

A live problem-scenario session makes a ready-made baseline hour: the reasoning is done by the people in the room.

Start this week

Start the reach log tomorrow and do your first baseline hour this month. In three months you will see both ends of the loop in your own numbers.

— Raghu Bharadwaj

Was this worth your time?
RB
Raghu Bharadwaj

Founder, TECH VEDA — 20+ years teaching the Linux kernel, device drivers and embedded systems.

Follow on LinkedIn

Get new posts by email

Kernel, embedded Linux and AI-era engineering — a few sharp reads a month. No spam.

We email occasionally and never share your address.