I’ve been turning over this distinction for months and I want to try to articulate it more clearly.
There’s a version of AI use that’s genuinely cognitively enhancing. I use it myself. I’ll take a draft argument, paste it in, and ask where the reasoning is weakest. I’ll ask it to generate the strongest objection to a position I’m developing. I’ll use it to check whether a concept I think I understand holds up when I try to explain it simply. These uses make me think harder. They extend the reach of my own thinking.
Then there’s another version. I’ve watched students, and sometimes colleagues, use AI to avoid the discomfort of not knowing something yet. You skip the productive confusion phase. You skip the period of uncertainty where the thinking actually happens. You receive a confident-sounding answer and move on without having built the neural scaffolding that a hard-worked understanding provides.
The second kind of use is insidious because it looks the same from the outside. The output might even be better. But something is lost in the process that doesn’t show up until later - usually in a conversation where someone has to reason through a novel problem and can’t, because they’ve been pattern-matching against AI outputs rather than developing the underlying capacity.
I see this in students who can submit competent essays but can’t explain their own argument in conversation. The essay exists. The thinking that produced it doesn’t.
I’m not sure this is unique to AI. Calculators, SparkNotes, tutors who give answers rather than guidance - all of these can be used to shortcut understanding. But AI makes the shortcut smoother, more comprehensive, and more convincing than anything before it.
I don’t have a policy solution. I’m more interested in whether people have found practices - for themselves or for others they work with - that preserve the thinking while using the tools.
The “productive confusion” framing is the one I keep coming back to.
There’s a specific kind of discomfort when you’re trying to work out something you don’t understand yet. It’s unpleasant in the moment and it’s where most of the actual learning happens. If you can skip it, most people will. I skip it sometimes and I know better.
The practice I’ve tried to build is using AI only after I’ve attempted something myself. Draft an outline first, then ask AI to find weaknesses. Write an explanation first, then compare it to what AI would say. The comparison is only useful if I have something of my own to compare against. If I start with AI, I’m just editing someone else’s work and calling it understanding.
Okay I’m going to push back a little here.
The “you’re not really thinking” argument has been made about every tool. Spell check, calculators, GPS. Nobody argues anymore that relying on GPS makes you a worse navigator, because navigation isn’t the skill that matters. Getting to the destination is.
The question is whether the thinking being skipped was actually important or whether it was just the tax you paid to get to the important part. For writing an essay about a book I don’t care about, I’d say the thinking isn’t that important. For understanding calculus, maybe different.
I’m not saying all thinking shortcuts are fine. I’m saying the answer depends on what the thinking was for.
The GPS analogy comes up a lot and I think it’s partially right and partially wrong.
Right: some skills that we used to consider essential have been genuinely superseded. Map reading is mostly irrelevant now. That’s fine.
Wrong: the spatial reasoning that navigation develops isn’t just about navigation. There’s evidence that people who navigate without GPS maintain better spatial memory and planning capacity in general. The skill is load-bearing for broader cognitive architecture in a way that skill itself doesn’t reveal.
The question for writing and structured thinking is whether those capacities are similarly load-bearing. Whether developing the ability to construct an argument from scratch, in your own words, with the discomfort of not knowing if it works yet - whether that process builds something that transfers to other domains. I think it does. The evidence for that is stronger than most people assume.
In CS the version of this I care about is whether you understand what the AI generated code actually does.
I use AI for boilerplate and I don’t think that’s cognitively harmful. I don’t need to deeply understand every CRUD endpoint I write. But when I use AI for something in a domain I don’t understand well, I’ve learned to force myself to trace through it line by line before I trust it. Not because the AI is usually wrong - it usually isn’t - but because if something breaks at 2am, I need to know the system well enough to debug it.
The “use it as a starting point, not an ending point” rule works for me. I don’t always follow it. When I don’t, it catches up with me eventually.
The version that worries me most in writing is the loss of tolerance for difficulty.
Good sentences are often the result of many failed sentences. The process of writing something badly and recognizing why it’s bad is where you develop taste. If AI is always providing the competent version before you’ve had a chance to attempt your own, you never develop the ability to distinguish competent from good, because you haven’t failed enough to know what failure looks like.
I use AI in drafting and I’ve noticed that my tolerance for sitting with bad sentences has decreased. I reach for the tool faster than I used to. Whether that’s harming my writing long-term I honestly can’t tell yet. But I notice it and I’m paying attention to it.