Every few weeks, another headline appears warning that AI is making us stupid. Students are apparently outsourcing their thinking. Writing is dying. Schools are banning AI tools while other schools are rushing to teach them. Somewhere, somebody is confidently announcing that an entire generation is about to lose the ability to think for itself.
I use these supposedly mind-rotting tools almost every working day, and every time I read one of these warnings I get the same feeling: we’ve been through almost exactly this before.
In the mid-1970s, pocket calculators went from expensive professional equipment to something ordinary people could afford to carry around. The reaction was remarkably familiar. Teachers worried that students would stop learning arithmetic. Exam boards banned calculators. Parents wondered whether children raised on push-button maths would become dependent on machines for things as basic as long division. For a while, the calculator was treated less as a useful tool than as a threat to thinking itself.
That sounds ridiculous now, and I think much of what we’re saying about generative AI will sound just as ridiculous fifty years from now.
I’ve written before about my grandfather, who gave me my first computer, and something about his career keeps coming back to me whenever this argument comes up. He lived through this transition professionally. The slide rule gave way to the calculator. The calculator gave way to the computer. Later, engineering software took over calculations that would once have consumed hours or days.
None of that made him less of an engineer, because the difficult part of engineering was never pressing the buttons. The difficult part was knowing which calculation mattered, what assumptions went into it, whether the answer looked plausible and what to do with it afterwards. The machine replaced the mechanism. It didn’t replace the judgement.
Fifty years after the calculator panic, mathematicians haven’t disappeared, and science didn’t grind to a halt because people stopped doing tedious arithmetic by hand. What disappeared was a lot of drudgery.
Long division became less important as a practical skill, in much the same way that lighting a fire with flint became less useful once matches existed. The human skill moved up a level. Instead of spending all our effort performing calculations, we could spend more time deciding which calculations were worth performing.
That’s how I think about generative AI. It’s a calculator for words.
Give it the mechanical parts of language and it can do an extraordinary amount of useful work: boilerplate, first drafts, summaries, rewording, translation, turning rough notes into something coherent, extracting structure from forty pages of prose, or writing the migration script you’ve already written twelve times before. What it can’t do is decide what’s worth saying.
It doesn’t know who you’re really writing for, or which trade-off you should make. It can’t tell whether a technically correct answer is sensible in the messy reality of your business, codebase or customers. That part is still yours, and we confuse the two constantly.
We see a machine produce fluent language and assume the machine has taken over the thinking. But fluency and judgement aren’t the same thing, and that becomes clearer when you look at how these systems actually work.
A large language model is trained to predict what comes next in a sequence. It does that again and again across an enormous amount of text until it becomes astonishingly good at producing language that fits the patterns it has learned. The effects are sometimes much stranger and more capable than that description makes them sound. We still don’t fully understand everything happening inside these systems, particularly where apparently simple prediction begins producing behaviour that looks like reasoning. That uncertainty only makes the operator more important.
I use these models every day, and what they’ve removed from my work is mostly the boring stuff: a regular expression I’d otherwise spend twenty minutes reconstructing from half-remembered syntax, a migration script following a pattern I’ve used dozens of times, the first version of some repetitive glue code, or a summary of a meeting I don’t want to replay in my head.
What AI hasn’t removed is any of the difficult work. I still have to decide what we should build and work out whether a system will survive poor network connectivity. I have to understand where something will fail, what data we might lose, what assumptions we’re making and who has to deal with the consequences when one of those assumptions turns out to be wrong.
Those are the parts of engineering that matter. My working day contains less drudgery than it did a few years ago, but it does not contain less thinking. If anything, the opposite has happened. Because the mechanical work is cheaper, a larger share of my time can now be spent on the things that actually require judgement.
The calculator comparison also explains something else people worry about constantly: AI getting things wrong. Calculators give wrong answers all the time. Put in the wrong number, put the decimal point in the wrong place, copy a figure from the wrong row of a spreadsheet or use the wrong formula, and the calculator will happily give you an answer.
It won’t flash a warning saying, “Karl, this looks suspicious.” It doesn’t know what you’re trying to do. Nobody considers that a profound failure of calculators. We understand that the machine performs the operation and the person using it is responsible for whether that operation makes sense.
Generative AI is not identical, of course, but the same basic discipline applies. Give a model a bad premise and it can produce several paragraphs of beautifully written nonsense built on top of it. Leave out an important constraint and it may confidently solve a different problem from the one you actually have.
None of this is a mysterious new category of human-machine interaction. Like any tool, it needs the person using it to understand enough about the problem to recognise when the output smells wrong.
We already know how to teach this. When people started using calculators, good maths teaching became more focused on number sense. You learn that twelve times nineteen should be somewhere around two hundred. So if your calculator says four thousand, you know something has gone wrong before you’ve even worked out exactly what.
That instinct matters more once the machine is doing the arithmetic for you. Terence Tao has raised a similar concern about AI, that these tools could deskill people the way calculators eroded number sense and GPS eroded spatial sense.
It’s a fair concern, and it argues for using the tool properly rather than against using it at all. What calculators taught us was that blindly trusting a machine is a bad idea.
Generative AI needs the same kind of instinct. You need to notice when a paragraph feels subtly wrong, recognise when an answer contradicts something you already understand, and know when a claim needs checking rather than copying straight into production or an email.
Call it text sense, if you like. The people who have it get enormous value from AI. The people who don’t are the ones most likely to be misled by it. That isn’t a new problem either. Tools have always rewarded understanding.
There’s another prerequisite that gets missed in most conversations about AI: you need to understand a problem well enough to describe it.
I was sitting with an engineer recently who was struggling with Gerrit after years of working mostly with GitHub. The underlying problem wasn’t really Gerrit. He didn’t fully understand the Git model underneath what Gerrit was doing.
GitHub can make it feel as though the same change can somehow exist independently in several places. Git itself is much more literal. Two commits with the same checksum are the same commit. In this case, Gerrit was rejecting a change because the relevant commit was already present. A local fast-forward had changed the history in a way he hadn’t noticed.
We eventually worked through it. What interested me afterwards was that AI wouldn’t magically have solved the situation for him. He couldn’t clearly describe what was happening because he didn’t yet have the mental model needed to describe it, and no machine could have handed him that. Even if a model had produced the correct explanation, he would have struggled to recognise why it was correct.
That’s the part of the AI story that usually gets told backwards.
If two people have access to exactly the same model, the person who understands the problem better will usually get more from it. They’ll ask better questions, provide better constraints, catch bad assumptions, recognise good answers faster and know when the model is confidently wandering off in the wrong direction.
The machine doesn’t flatten expertise. It amplifies the difference.
Calculators work exactly the same way. A calculator is incredibly useful to someone who understands mathematics. Give the same calculator to somebody who has no idea what numbers to enter or what operation they need, and it’s almost worthless. AI just makes that prerequisite harder to ignore.
There’s another reason I’m optimistic about all of this. When the mechanical parts of a craft become cheaper, people tend to become more ambitious.
Calculators expanded what ordinary engineers and scientists could realistically attempt, and computers took that much further. Calculations that once required teams of people became cheap enough to run repeatedly, simulations became practical, and engineers could explore more designs before committing to one.
Language will probably follow the same pattern. The developer who never had time to write proper documentation suddenly can. The small business owner who struggles to write confidently in a second language can communicate more clearly. The engineer can explore three possible implementations in the time it previously took to build one. The researcher can get through material that would otherwise have sat unread.
And yes, plenty of terrible AI-generated rubbish will be produced too. Every cheap technology increases the supply of both good work and bad work. That doesn’t bother me very much. What interests me is what happens when the cost of trying an idea drops dramatically, because people try more ideas.
This is where I part company most strongly with the way AI is currently discussed. Calling these systems “calculators for words” isn’t nearly dramatic enough for the AI industry.
A calculator for words is useful, possibly extremely useful, but it isn’t a trillion-pound story. A synthetic mind about to surpass humanity is a much better story. That story raises money, fills conferences, drives headlines and makes products feel historically inevitable rather than merely useful.
A surprising amount of the fear around AI comes from the same machinery that produces the hype. The companies selling these systems benefit when we think of them as something completely new and almost unknowable.
I wrote more about that in The AI apocalypse is a distraction, because the obsession with a hypothetical machine apocalypse has distracted us from much more immediate, ordinary problems.
There’s something commercially convenient about convincing everyone that this technology represents a clean break with everything that came before. I don’t think it does. We’ve seen this pattern repeatedly.
The calculator arrived and people worried that students would stop thinking. Then it was computers making workers dependent, search engines meaning nobody would remember anything, and GPS costing us our sense of direction.
Some of those concerns contained a real cost. People probably do remember fewer phone numbers. Many of us are worse at navigating unfamiliar cities without GPS. Plenty of people have weak mental arithmetic because they’ve always had a calculator nearby.
But that isn’t the whole story. We traded some mechanical skills for far greater capability, and we decided, mostly correctly, that the trade was worth it.
AI will have costs too. Some people will use it lazily, some schools will handle it badly and some companies will automate things they shouldn’t. Plenty of people will become dependent on it without ever developing the judgement needed to check its work.
Those are real problems, but “some people will use a tool badly” has never been a convincing reason to pretend the tool itself isn’t useful.
Generative AI is a remarkable, fallible calculator for words. That description sounds much less exciting than “artificial intelligence”, and I suspect that’s precisely why I prefer it. It puts the machine back in its proper place.
The machine does the mechanical work. You decide what matters, what’s true, whether the answer is any good and what happens next.
Calculators didn’t replace mathematicians. They made arithmetic cheap and judgement more valuable. Generative AI is doing the same thing to language, and I’d much rather spend our time working out what becomes possible when the boring parts get cheaper than pretending the calculator has suddenly become the mathematician.