I would like to open my computer one morning, ask what happened in the world, and hear: “Nothing particularly interesting for you today. Go for a walk with your girlfriend.”
I would have to teach it to do that. Explain what matters to me, correct it when it gets things wrong, and tell it when I have had enough. I would not get that assistant simply by installing whatever a company sells me. Even while preparing this essay, I had to keep correcting what the AI thought I wanted.
I was dictating thoughts in Polish for an essay I wanted to publish in English. I work in marketing, websites, and content, increasingly with AI tools that generate text and images. I also co-own a company that produces 3D imagery. I wanted to explore what happens as machines take over more of the production and more of the reading: the prospect of an internet largely occupied by software, and the appeal of receiving information through an interface shaped around me.
I asked the model to preserve my thoughts and find ideas I could set them against. It gave me a list. The first item connected what I had said to information growing faster than our capacity to absorb it, with a reference to Herbert Simon. The connection made sense. But it made the essay about something other than what I wanted to write about.
I replied: “That isn't my idea. Take it out.”
I wanted to talk about who controls the AI through which I encounter the world, and how much work I would have to put into making it useful to me. I had to explain that difference to the model before we could continue.
Over the past four years, I have watched AI take over simpler modeling and finishing work in 3D imagery. An architectural project still needs an architect, and a design still needs a designer. How many steps are we from replacing them too? I do not know. In building websites, I can already direct code, graphics, and revisions through the same conversation. But I still have to put creative work into it. Asking for “a ten-thousand-dollar website” tells the system very little. The price is not a creative brief.
The amount of thought involved and the amount of production labor involved are beginning to come apart. I can get an idea onto the screen quickly enough to try it while it is still fresh. Whether that leads to more thought depends on what I do with the result. In this essay, a coherent interpretation arrived before I had settled what I wanted to say. Accepting it would have made the work faster. Disagreeing with it made the work mine.
Paul Graham's “Writes and Write-Nots” takes that danger seriously. He expects AI to erode the middle ground of writing ability, leaving accomplished writers and people who no longer write. His concern goes beyond the value of practice: he argues that some kinds of thinking can happen only through writing.
I agree that good writing is difficult, and language models are already remarkably capable at fluent, serviceable prose. They learned from human writing, including work much better than most of us produce. When a machine can formulate a sentence for you, the pressure to learn how to formulate it yourself weakens.
We also think through drawing, reflection, and conversation, although that alone does not answer Graham's argument. What I want to know is whether working on a text with AI can involve the effort he associates with writing. That is what I was doing when I read its interpretation, worked out why it was wrong, and explained what I meant. I was working on the thought as well as the wording.
I do not know what I would have discovered by writing every sentence myself. I may have missed something by working this way. But I do see a difference between accepting a finished text and going through it to check whether I agree with what it says. I expect a division between people who use AI consciously and people who use it to avoid thinking. Both groups will use it, and their texts may look equally good.
The stakes extend beyond writing. Machines already generate material that other machines retrieve, read, and summarize. A publishing system whose authors and readers are mostly software feels dispiriting to me. Yet I would readily ask an assistant to bring me information rather than visit every website myself. I am helping create the conditions for a future that troubles me.
I cannot remember the last time I opened a CMS admin panel, even though I make changes to websites every day. I can imagine reading in that same conversational interface, with AI selecting material from the web for me. With this essay, I could see an interpretation and reject it. If an assistant selects what I see all day, I will also need a way to find out what it has left out.
Perhaps we are living through the worst phase of AI right now.
We already encounter it in Google, Facebook, TikTok, and YouTube. It selects what we see, and we have little control over the goals behind that selection. We can signal what we like, but those signals go into systems run for the benefit of the companies that own them. If something keeps me browsing, that can be a good result for the company even when I would rather have spent my time elsewhere. The AI is already working. I cannot simply tell it to work according to rules I have chosen.
That is why I can imagine more AI giving us more control than we have now. I want to be able to tell an assistant what information to look for, why a recommendation was wrong, and when to stop. I want those instructions to govern what it does. If that becomes possible, today's experience may look like a period in which we were subject to these systems without having much say in how they treated us. I do not know whether we will get there. Better models alone will not make it happen.
An assistant running on my computer appeals to me, but its location is not enough. I need to be able to change the rules for selecting information, ask why it proposed something, and revisit what it omitted. When I tell it to stop, it should stop. Those are things I would want to check before calling it mine.
I would also have to spend time teaching it. A company could supply the model; it could not do all that work for me. I would correct its choices, explain why something engaging was nevertheless unwanted, and teach it when a subject deserved more depth. It could help with information coming in and material going out, while I remained able to change its decisions. That is the assistant I mean when I talk about a computer suggesting a walk.
How many people will want to do that? It is easier to take something ready-made. TikTok or YouTube offers something immediately, while teaching an assistant requires me to work out what I want and keep correcting it. Thinking takes effort. For plenty of people, the ready-made system will remain more attractive, even if they could have more control.
I am more hopeful when I think about how little technical knowledge this conversation required. I spoke in Polish, read the proposed interpretation, and explained what was wrong. I did not need to find a setting or learn a command. I can imagine an interface showing my original thought beside the proposed text so I can point to the difference and work on it. More people might be willing to teach their AI that way.
I do not believe active use has to remain limited to a small group. Someone who struggles to write could still develop an idea through conversation, check what the system makes of it, and keep changing it. That takes effort too, but it is an effort more people may be willing to make.
I still want the computer that suggests a walk. But there is a difference between deciding that nothing is worth reading and recognizing that I have read enough for today. I would want to know which judgment it had made, and be able to disagree. Then I might close the computer.