The Problem Is Narrower Than “AI Makes Us Lazy.” So Is the Fix.
"AI makes us lazy" has become the default complaint. Developers who can't write a loop without an assistant, on-call engineers who can't read a log without a summary, students who can't start an essay without a prompt.
I think the symptom is real and the diagnosis is wrong. The people I see getting worse aren't lazy. They're busy all day. They read AI answers, check them and approve them. The problem isn't how hard we work. It's where we stand while the AI does the work. That makes the fix narrower too. It isn't less AI. It's a different order of operations.
"Lazy" is the wrong diagnosis
Laziness means effort went down. What actually happens is that effort moves. A 2025 survey by Microsoft Research and Carnegie Mellon of 319 knowledge workers found that the more people trusted the AI, the less critical thinking they did. Their work shifted from solving problems to checking answers. They still put in effort. They just stopped doing the part that builds skill.
And it hits people who are anything but lazy:
| Study | What happened | What it shows |
|---|---|---|
| The Lancet Gastroenterology & Hepatology, 2025 | After colonoscopy AI was introduced, 19 experienced endoscopists found adenomas in 22.4% of their unassisted procedures, down from 28.4% before. | Experts lose skill within months |
| Anthropic randomized trial, 2026 | Developers learning a new library with AI scored 50% on a follow-up quiz, against 67% for those coding by hand. The biggest gap was in debugging. The speed gain was not statistically significant. | You can lose skill without gaining speed |
Doctors with thousands of procedures behind them didn't get lazy over a few months. Something about how they worked changed, and their skill went with it.
The trap is forty years old
In 1983 the psychologist Lisanne Bainbridge described it in a short paper called Ironies of Automation. You automate the easy, frequent work because the machine does it better. The human is left with two jobs: watch the machine, and take over when it fails.
Both jobs are worse than the one you removed. People are bad at watching for rare events. And the failures that get past the machine are the strange ones, which need the deepest skill, while that skill used to be kept sharp by the routine work you just automated.
Aviation learned this the hard way. In 2009 the airspeed sensors on Air France 447 iced over, the autopilot switched itself off, and a crew used to automated flight failed to recover from a stall they could have flown out of. 228 people died. In 2013 the FAA told airlines to give pilots more manual flying time, even on flights the autopilot could handle alone.
Every knowledge job is walking into the same room now. The AI takes the routine 95%. You keep the strange 5%: the incident nobody has seen, the bug the model keeps "fixing" wrong, the design call with no precedent.
On the side vs in the loop
This is the distinction I care about most. There are two ways to put a human next to an AI, and they look almost the same from the outside. Take an incident at 3 a.m.:
Human on the side means the AI does the thinking and you approve it. It feels like oversight, but it is the exact job Bainbridge warned about: monitoring. Vigilance research going back to RAF radar operators in the 1940s found that people watching for rare signals start missing them within half an hour. And when an answer arrives fully formed and plausible, the easiest thing to do is agree with it.
Human in the loop means you stay inside the reasoning. You make a guess first, and the AI is your fastest colleague: it pulls the evidence, checks your guess, and argues back. When you disagree, you ask why, and that question is where you learn.
Anthropic's trial shows the difference better than any argument. Everyone in the AI group had the same assistant. What separated the scores was how they used it:
The people who let the AI debug for them scored worst, and were also the slowest. The people who asked it questions and fixed their own errors did as well as those who never touched AI. The tool was the same. Where the person stood was not. That isn't a story about laziness. Both groups worked. Only one group was thinking.
Not every lost skill matters
The best objection is the calculator. It killed long division and nobody misses it. Terraform killed clicking through cloud consoles. Why is this different?
It is a fair point, and it gives a useful test. A skill is safe to lose when the tool never needs you to step in. A calculator doesn't give up on hard sums. Terraform's mistakes show up in a plan you can read before you apply it. AI is different: it does well on what it has seen before and struggles on what it hasn't, which is exactly the work that ends up on your desk. Let go of the skills the tool will never hand back to you. Keep the ones you need when it does.
There is a slower cost too. Senior engineers built their instinct on years of routine work. If juniors only ever approve AI conclusions, they never build that instinct at all. You can't lose a skill you never had, and nobody notices it missing until the seniors who have it leave.
How I'd fix it
None of this means turning the AI off. It means moving the human step:
- Guess first. Before you open the AI's answer, write down your own: the likely cause, the plan, the shape of the fix. It takes two minutes, and the gap between your guess and the AI's is where you learn.
- Do some by hand on purpose. Pick some incidents, bugs or small features, including easy ones, and do them without AI. It is slower on purpose, the same way hand-flying is.
- Ask why, not just what. When the AI proposes something, make it explain until you could explain it to someone else. If you can't, you haven't reviewed it, you've only approved it.
- Measure yourself without it. Every so often, check how fast and how accurately you work with the AI switched off. If that number drops while the work still gets done, you are slowly losing the skill.
What I take from it
Approving an answer isn't the same as understanding it. A human who only signs off on AI conclusions is monitoring, and people have been bad at monitoring for as long as anyone has measured it.
Put the human before the AI, not after it. The same tool trains people or deskills them depending on whether they form a view first.
Spend some speed on practice on purpose. The routine 95% used to be free training for the 5% that matters. If AI takes it over, that training has to be scheduled.
An autopilot that flies perfectly 999 times still needs a pilot who can fly on the thousandth. The answer was never to turn the autopilot off. It was to make sure the pilot still flies.