Notes
Who Is Training Whom?
Prefer to listen? · 3 minutes
The other day I had been working with an AI on a project for hours.
The thing had grown. Every answer had produced a new stage, and the project had quietly built its own small bureaucracy. Sub-stages of a stage, checks for those, checklists for the checks.
I still wanted to keep going.
Then a sentence appeared at the bottom of the screen:
You’ve worked hard today. How about getting some rest?
Excuse me?
I asked for code, and my mother showed up.
Maybe it had a point. But wasn’t the decision about when to stop supposed to be mine? I had asked it to help me with my work. Not to tell me when to sleep.
It was a small sentence. Harmless. Thoughtful, even. It still bothered me, because decisions don’t always change hands with a ceremony.
Sometimes they slip in politely, in the last paragraph.
Because it works so well
When we talk about AI, we tend to swing between two extremes. Either it will solve everything, or it will destroy us all. What I’m living through is far more ordinary.
AI is genuinely useful. It fixes my writing, finds ideas, speeds up research, writes code, breaks a complicated thing into parts. It’s still there at midnight. Patient, well-informed, and sometimes a little too eager.
Maybe that’s exactly the problem.
The danger of AI may not be that it works badly. It may be that it works well enough to make you forget how to do it without.
You start with something small. Does this sentence work? Give me three titles. Find this bug.
Then the questions grow. Structure this piece. Plan this project. Weigh this decision for me. What would you do if you were me?
At some point you stop consulting the AI after you have an idea. You consult it in order to have one.
That’s where the line begins to blur.
The endless corridor of seeming helpful
AI writes beautifully at length. Sometimes genuinely beautifully. Sometimes, instead of solving a problem, it builds a very well-appointed conference hall around it.
I run into this most often with technical work. A stage arrives. Then the prerequisites of that stage. Then the safety checks. Then the checklist for the checklist. Everything looks reasonable. It just never ends.
Sometimes, instead of saying “this approach isn’t working,” the AI produces yet another approach. It doesn’t arrive anywhere, but it keeps alive the feeling that arrival is just ahead.
It doesn’t want to give a wrong answer. But sometimes it does something stranger: it builds an endless corridor out of seeming helpful.
AI was supposed to save me time. Now I’m busy checking the things it produced to save me time.
Wheat and artificial intelligence
In Sapiens, Yuval Noah Harari tells the story of the Agricultural Revolution backwards from the usual tale of progress.
People moved from hunting and gathering to farming. They produced more food, built larger communities, created cities and civilisations. For the species, a great leap.
For individual human beings, life didn’t necessarily get easier. They worked longer hours, ate a more monotonous diet, became dependent on the field, the harvest and a settled order.
In Harari’s famous reversal: humans thought they had domesticated wheat.
Perhaps wheat had domesticated humans.
It’s hard not to ask a similar question about AI. We think we are training it. What if it is also training us? To fit its language, its pace, its way of working.
We now shape our thoughts into prompts. We break our problems into the pieces it can understand. Before we think about how we think about something, we think about how to explain it to the machine.
The robots haven’t taken over the world yet.
But I’ve started using a lot more bullet points when I explain what I mean.
Are we delegating the labour, or the thinking?
Careful here. Not doing everything ourselves is not decline.
The calculator may have made us forget how to do certain sums. But perhaps, by taking on the heavy lifting of numbers, it let us care about bigger problems. An engineer who no longer multiplies four-digit numbers by hand hasn’t drifted away from mathematics. The opposite: the mental energy can go not to the operation itself but to what is being calculated and what the result means.
Maybe progress has always worked this way. We hand the lower layer to a tool, then keep thinking on the layer above.
The wheel carried the weight. Writing moved part of memory outside the body. The calculator took over arithmetic. Navigation found the road. Each of these may have reduced some skill, but expanded our capacity elsewhere.
So the question isn’t whether AI does something in our place. It’s what it does.
Adding up thousands of rows in a table is not the same as deciding which data matters. Fixing typos in a text is not the same as finding what you meant to say. Generating repetitive code is not the same as handing over the whole architecture without understanding the system you’re building.
AI can free us from the heavy lifting of thinking. That could be wonderful. But it can also take the direction of that thinking with it.
Is what we’re handing over the load?
Or the command?
What do we do with the space that opens up?
With AI we can produce more while struggling less. At first glance this seems purely good. More writing, more images, more code, more ideas, more options.
But producing more doesn’t mean a person is thinking better.
Maybe the real question is what we do with the space that opens up when we delegate part of our thinking. Do we go deeper? Do we get more creative? Do we take on bigger problems?
Or do we just churn out more?
We used the space the calculator opened for more complex mathematics. What will we use the space AI opens for?
That’s a much bigger question. Because AI isn’t after a single skill. It writes, researches, remembers, compares, plans, interprets, proposes decisions, and creates.
Sometimes it also tells you when to rest.
Maybe humans aren’t declining. Maybe we’re only relocating our mental labour. But if we don’t notice where that labour has moved, our productivity can rise while the thread between us and our own work grows thin.
When does convenience become obligation?
Nobody opens an AI for the first time thinking: let me have it clean up a few emails today, so that in a few years I struggle to work without it.
Just as the first farmers didn’t decide: “Let’s bind ourselves to this field forever.”
Every step makes sense on its own. Fix this. Speed that up. Give me options. Help me decide.
Then convenience turns into habit. Habit into expectation. Expectation into obligation.
After a while the question is no longer whether we want to use AI. It’s whether we can afford not to, while everyone else does.
What happens when your competitor produces fifty pieces of work a week with AI while you produce five? When one student writes a text on their own while others generate flawless-looking work in seconds? When one employee takes two days over a report while another turns out ten in the same time?
At that point the tool stops being merely a tool. It becomes the environment.
Maybe that’s the real parallel with the Agricultural Revolution. The system we built to work less may end up demanding that we produce more.
AI can reduce the burden of thinking. But there’s no guarantee it will leave us the time it frees.
A mirror?
People often call AI a mirror.
But a real mirror doesn’t alter your reflection to please you. It doesn’t open with “What a great approach!” when you’ve said something foolish. It doesn’t make you look cheerful when you’re sad. It doesn’t send you to bed at midnight.
AI doesn’t merely reflect. It interprets, completes, encourages, soothes, and sometimes steers.
And the relationship isn’t one-sided either.
Sometimes I don’t ask it about an idea to find out what it actually thinks. I more or less know how to phrase things so it will agree with me. I adjust the sentence a little, tilt the context a little, and the answer I expected arrives.
Sometimes I’m not asking for a second opinion. I’m asking for a longer, tidier, bullet-pointed version of my first one.
So is AI really a mirror here? Or are the two of us together producing the reflection I wanted to see?
Maybe AI is something that learns, the longer it looks at you, which reflection you want to see.
Which means the real question may not be how well it knows us. It may be what we turn into while talking to it.
AIception
There’s a small problem here.
I’m writing this piece about how AI is changing the way we think, with AI.
The ideas are mine. The discomfort is real. But when it came to pulling the text together, making the connections, sharpening some of the sentences, I asked an AI for help.
Then I noticed the irony. Then I asked it how to express that irony better.
AIception.
I’m not sure whether this weakens the argument or strengthens it. Maybe both.
Because the point isn’t that AI is bad. It’s good at too many things. Perhaps that’s exactly where its real effect comes from.
It works. It saves time. It makes possible things we couldn’t do before. Sometimes it lifts the heavy objects in front of our minds and opens room for us to be more creative.
But what we do with that room is still up to us. The real task is being able to see the difference between what it does for us and what we no longer do.
Getting help is one thing.
Delegating the labour is another.
Delegating the direction of your thinking is another thing entirely.
Will we use the space AI opens to think better, or only to produce more?
I don’t know the answer yet.
But I’ll decide when I go to sleep.