I have written probably a thousand texts in my life. Each one left something in me that had not been there, or had not been clear, before I started. As a child, writing was already the most effective way I had to learn, for exactly this reason, though I could not have named the reason at the time. Sitting down with a pen and working out what I had actually meant to say was how the learning happened. I found the reason later, in writing workshops I attended as an adult, where the same question kept coming back: does writing record a thought, or does it make one?
The pitch for AI writing tools names itself plainly. It is a “writing assistant”, there to handle “phrasing, structuring and especially grammar”. The division sounds tidy: the machine only handles the sentence, once the author has already supplied the thought.
In my experience, a thought takes its shape at the moment I write the sentence that states it. The sentence I start rarely ends where I expected, because at the start there was often no plan at all. One word opens a direction that was not visible a moment earlier. Across a whole text the order can settle out differently once the argument is built, but at the level of a single sentence or paragraph, this happens even in the most matter-of-fact writing.
Anyone who writes recognises the other case too. I sit down with a claim, and three paragraphs later I notice it is wrong, or it has stopped mattering. There is no one in the room to object. The claim meets its first real resistance when I try to put it into a sentence that has to be exact. That precision does its work even when it leaves gaps.
Handing the sentence to an AI skips exactly that moment. I type in a vague thought and get back a precisely worded text. The assumption that the thought was already clear from the start overwrites the vagueness. The system may never have thought it through at all. It only formulated the thought as though it had, with no way of noticing whether the result matches what I actually meant.
I tested this on myself, out of a fascination I want to record honestly: the idea that I might not need to think at all, because the model would do it for me. I took a thought I had jotted down as a raw prompt and asked ChatGPT to write it out in full. What came back read as though I had written it myself. It did not hit what I meant. It said something similar, plausible and pointing in the right direction, but it missed the point itself. Recognising that would have meant trying two wrong formulations and working out why each one was wrong. A model cannot do that.
Philosophers write their own sentences for a specific reason, rather than dictating the argument or having someone else summarise the text. Ludwig Wittgenstein spent years rearranging the propositions of the Tractatus Logico-Philosophicus (1921) until the sequence matched the argument he wanted to make. The order itself was the thought. The book opens: “The world is everything that is the case.” Reverse the two halves of that sentence, “Everything that is the case is the world”, and the emphasis falls somewhere else entirely. Wittgenstein was doing philosophy at the level of word order.
Letting an AI supply the sentences treats the force of a thought as something thin. It assumes content and form are separate, so the form can be handed off on its own. That split does not exist in how thinking and writing actually happen. Content takes shape inside the form. When I hand off the formulating, the thinking that was meant to happen inside it never occurs.
I do use AI for structuring, and it is a useful tool for laying out a logical outline. Structure in an argument, though, is the logic that makes the argument work. The sequence in which ideas appear decides what they end up meaning. Run in a different order, the same argument means something else. Handing over the structure is handing over the argument itself, and with it the decision about what the finished text actually does.
A consultant who has a model phrase the final recommendation is handing over the recommendation itself. By then it is the model’s.
Most people I know treat AI as a writing assistant, nothing more. Writing is one of the last places where I still catch myself thinking in real time, sentence by sentence. A model can produce a paragraph that reads as though someone had worked through a question and arrived somewhere. The paragraph looks like evidence that someone thought it through. No thinking took place.