Do It by Hand First: Method for Keeping Critical Thinking Alive in the Age of AI

The first time I hit a genuinely hard problem at work, I solve the whole thing by hand.

I'll use every tool I have to research it: documentation, search, and yes, AI. But I make myself take each step. I want to feel where it's hard and where I was wrong. When it's finally working, I stop and write down how I actually got there, while it's fresh.

The second time the same problem shows up, I don't do it all by hand again. I take those notes and turn the boring, repeatable parts into a script (usually a small computer program), so I can spend my attention on the one or two decisions that actually need judgment.

The third time, or when I know I'll face it forever, I build something that handles it automatically. By then I understand the problem well enough to trust a machine with it. Just as importantly, I understand it well enough to notice the day the machine gets it wrong.

For years I thought this was a method for going faster. It isn't, really. It's a method for making sure that when I finally hand a task to a machine, I've already earned the understanding to catch it when it fails. And that turns out to be the exact thing AI makes it very, very easy to skip. For our kids, and for us.

This post is part method, part evidence, and part opinion, and I'll tell you which is which as we go. It is not an argument against AI; I use these tools every day, and this whole site exists to help you use them too. It's also not a research paper. It's one practical way to think about the single skill I most want my kids to keep: the ability to actually think, even when a machine offers to do it for them.

A method I borrowed from engineering

Strip the jargon off what I described above and it's a simple ladder with four rungs:

Do it by hand. The first time, take every step yourself. Use tools to research, but do the actual work.

Write it down. Once it works, capture how you got there, the wrong turns included.

Script the parts that repeat. The second time, automate the routine steps so you can focus your attention on the parts that need real judgment.

Automate it fully, but only once you've earned it. Hand the whole thing to a machine only after you understand it well enough to check its work.

The important part isn't any single rung. It's the order, and the fact that automation comes last. You don't start by handing the problem to the machine. You end there, and you only trust the handoff because you built the understanding first.

That order is the whole argument of this post. The tempting thing about AI, for a busy adult and for a twelve-year-old with homework due, is that it lets you jump straight to rung four on a problem you've never once done by hand.

Why the order matters more than the speed

Here's the part that's actually well-studied, and it surprised me when I first read it: the struggle isn't a cost you pay to get to the understanding. The struggle is how the understanding forms.

Three findings from decades of learning research point the same way.

You remember what you produce better than what you're handed. Psychologists call it the generation effect: people who generate an answer themselves, even a partial or wrong one, remember it far better than people who simply read the correct answer.1 The act of producing it is what makes it stick.

Retrieving something is a stronger teacher than reviewing it. Testing yourself on material, pulling it out of your own head, produces much more durable learning than reading it again, even though re-reading feels more productive.2 The effort of retrieval is the point.

Learning that feels harder now tends to last longer. Researchers call these desirable difficulties: conditions that slow you down and make you work often produce deeper, more durable understanding than conditions that make things feel smooth and easy.3 Easy and fast frequently means shallow and gone by next week.

I want to be careful here, because this is exactly where it's easy to overclaim. So let me under-claim instead. None of these studies measured "critical thinking" directly; they measured memory and learning. And I know of no study that has tested whether my particular four-rung method builds it. What the research does support is narrower and sturdier: understanding you build yourself sticks better than understanding you're handed.

Critical thinking leans on that, though it isn't the whole story. You can question a claim in a field you barely know, asking who's behind it, what they gain, and whether other sources agree. That's the bias-spotting habit worth teaching all on its own. But the deepest kind of check, knowing an answer is wrong because you understand the thing, only comes from having done the work. Skip the doing, and you're left with just the surface checks.

That's rung one and rung two of the ladder, and it's why they come first: do it by hand, and write it down, because that's the part that's building you.

What happens when you skip straight to the answer

Now the other side: what the research says happens when you routinely hand the thinking off before you've built it yourself.

We remember where to find things instead of the things. In a set of well-known experiments, people who expected a fact to stay available to them later remembered the fact itself less well, but remembered where to look it up better.4 This is sometimes called the "Google effect," and on its own it's not a disaster; nobody memorizes phone numbers anymore and we're fine. The trouble starts when what you've offloaded isn't a phone number but the reasoning itself.

We over-trust automation and stop catching its mistakes. Decades of research on people working alongside automated systems keeps finding the same pattern: when a machine usually gets it right, people stop watching closely and miss the errors it does make.5 And a skill you stop practicing fades. You can see a version of it in studies that link heavy reliance on turn-by-turn GPS to weaker spatial memory over time.6 The skill you don't use is the skill you lose.

And the newest research points the same direction for AI specifically. A 2025 study of knowledge workers by researchers at Microsoft and Carnegie Mellon found that the more people trusted an AI tool, the less critical thinking they reported doing. They shifted from doing the work to checking the machine's output, and sometimes skipped even that.7 A separate survey of 666 people found a similar link between frequent AI use, "cognitive offloading" (letting a tool hold the thinking for you), and lower critical-thinking scores.8

I'll be honest about how much weight those two can bear. Both are correlational, and both rest on people reporting on themselves. They can't prove which way the arrow points. It's entirely possible the causation runs backwards: the people who lean hardest on AI may already be the ones least inclined to dig in. Hold them as a strong hint, not a verdict.

You may also have seen scarier headlines about an MIT experiment that wired people up to EEG sensors while they wrote essays with ChatGPT and found weaker brain activity, the so-called "cognitive debt" study.9 I'm raising it because you've probably seen it, not because you should lean on it: it's a small, not-yet-peer-reviewed preprint that has already drawn a formal methodological critique. It's interesting; it isn't proof. And it's a tidy example of the exact habit this whole post is about. Don't bank a confident claim just because it flatters what you already suspect.

None of this says AI makes you stupid. But the honest, older, sturdier learning research and these newer hints rhyme with each other, and they rhyme with the thing this whole series keeps coming back to: AI is confident whether or not it's right, and if you've handed it the thinking, you have nothing left to catch it with.

The one rule that makes automation safe

So what's the difference between my engineering ladder and just... using AI to skip the work? On the surface they look similar. Both end with a machine doing the task.

The difference is one rule: you earn the automation, and you never fully take your hands off the wheel.

We already do this with a tool nobody panics about anymore. We let kids use calculators after they understand what multiplication is, not instead of learning it. The calculator is rung four. Arithmetic by hand is rung one. We got the order right on that one, mostly by accident and a lot of argument, and it works. A kid who understands multiplication can use a calculator all day and still notice when the screen says something absurd. A kid who never learned can't.

You could push back here, and thoughtful people do: maybe AI is simply the next calculator, and in twenty years worrying about "cognitive debt" will sound as quaint as the old panic that calculators would ruin arithmetic. That might turn out to be right for some tasks. I hold this whole argument a little loosely.

But there's a difference of scale. A calculator does one narrow thing, and you can spot-check its answer at a glance. A general-purpose AI will draft the essay, build the argument, and decide what matters. That's the whole chain of reasoning, not one arithmetic step, and it arrives in fluent, confident prose that's much harder to eyeball for "absurd." The wider the task you hand off, the more you needed to have done it yourself to judge the result. That's not an argument against the tool. It's an argument about the order.

That "still notice when it's absurd" is the part that matters, and it's a habit, not a one-time achievement. In the last post, I had an AI build a real piece of software in an evening, and the single most valuable thing I did was refuse to take its word that the work was good. I made it prove it, against a check the AI didn't control, and I kept a manual review step for myself on the one thing I actually knew how to judge. That's rung four done right: automate the work, keep your hand on the check.

The wrong version, the one to actually worry about, is handing the whole thing over on a problem you've never done, then trusting the output because it sounds confident. That's not using a tool. That's skipping straight to the top of a ladder you never climbed.

How to raise a kid who earns the shortcut

Picture the moment this actually happens. Your kid is at the kitchen table with a socials assignment: a paragraph on why the Fraser River mattered to the communities that lived along it. You watch their thumb drift toward the phone before they've written a single word. That drift is the whole ballgame.

Here's how I try to turn the ladder into something you can do in that exact moment. None of it requires you to be technical. And it scales with age: a six-year-old "doing it by hand" is sounding out a word; a sixteen-year-old's is sketching the argument before they write it. Match the ambition to the kid. An upcoming age-by-age post goes deeper on that.

Let them do it by hand first, and let it be a little hard. When a real problem shows up, whether it's that Fraser River paragraph, a tricky math question, or a text to a friend they're nervous about, the instinct is to relieve the struggle fast. Resist it for one round. The difficulty isn't the enemy of learning; it's the engine of it. Let them produce a messy first attempt before any machine touches it.

Make them show their thinking, out loud. Not "what's the answer" but "how did you get there, and where did you get stuck?" This is rung two, writing it down, in conversation form. It's also the single best defense against outsourced thinking, because a kid who can narrate their reasoning has actually done some.

Introduce AI as the second pass, not the first draft. This is the one habit I'd most want you to take from this post. Flip the usual order: do the work first, then bring in the AI to critique it, poke holes in it, or offer a version to compare against. "Here's my answer; what did I miss?" builds judgment. "Gi

[truncated for AI cost control]