Yes, I use AI

I use AI to write. I use it to research, to draft, to code, and to argue with. This essay included.

I expect that within a few years the disclosure will read oddly, the way it would read oddly if I told you I had used a pen. Or a spreadsheet. Or a search engine. Tools stop being remarkable once everyone has them, and the announcement stops being a confession and becomes a tic. We are not there yet, so here is the confession.

What I still own

I publish under my name. That settles most of the question.

Mark Graph is a pen name, not a disguise. Every page on this site has my real name in the sidebar. What matters is not the legal name but the persistence. One byline carries everything I have written. Get something wrong and it stays there, next to whatever I publish next.

Every claim in what I write is mine to defend. If the analysis is wrong, I was wrong. Not the model. These are my thoughts, my reading of the evidence, my sense of how the world works. They include my mistakes and my errors of analysis, and there are plenty. AI has not removed them. It has changed which ones survive long enough to reach you, which is a real improvement, but it is not the same thing as being right.

What the tool does is help me express myself. It helps me anticipate the criticism I am going to get. It helps me calibrate a claim so that it is as strong as the evidence will carry, and no stronger. Those are all things I was trying to do before. I am now better at them.

Three ways I use it

The first is code. I once wrote all the Python myself: pull the series, clean it, reshape it, build the plot. I still know how to do that. I just no longer spend the afternoon on it. The thinking was never in the boilerplate. The thinking is in choosing what to plot, and in knowing what the series actually measures, and in noticing when a break in the data is a break in the world rather than a change in the collection method. None of that has moved.

The second is long-form writing, which I will come back to.

The third is reading. When someone on X makes an argument I do not follow, or one I think is wrong, I work through it with AI before I respond. Sometimes the argument turns out to be better than it looked and I learn something. Sometimes I find the flaw faster than I would have alone. Either way I reply having understood the thing I am replying to, which is not the norm on that platform.

The coding and the reading save me time. The writing changes what I publish.

Arguing with the machine

The drafting is the least interesting part. Everyone knows the machines can produce fluent text at speed, and summarise a literature faster than I can. The arguing is where the value sits.

I make a claim and I ask the model to take the other side, hard. Not to be agreeable, not to hedge, but to find the weakest link and pull on it. Then we go a few rounds.

What comes back is not a verdict. The model cannot tell me whether I am right. What it can do is generate objections on demand, immediately, in volume, and without the social cost of asking a colleague to attack your work for the fourth time in a week. An objection does not have to be good to be useful. A weak one still makes me say why it is weak, and I usually learn something from having to say it.

Sometimes I change my mind. I walk in believing something and walk out believing less, or believing something adjacent. Sometimes the model changes its. Who moved is not the interesting question (and if the model moved, I do not know whether my brilliance or its obsequiousness won the day). The draft that comes out is better than the draft that went in, because it has had to survive something, and that holds whichever way the argument went.

Often nothing moves, and I come away knowing which paragraphs are load bearing and which ones I was leaning on out of habit. A weak argument that reaches publication does far more damage than one killed on a Tuesday afternoon.

None of this is new. Good academics workshop papers before they present them. Barristers moot arguments before they walk into court. Politicians rehearse the interview before they give it. The practice is old and well understood. Until recently it needed other people, who had to be willing to give up an afternoon and then willing to do it again. A solo writer never had that. What changed is the price.

And by the time I publish I know how I would answer the criticisms I can foresee. Not word for word, but I have been through the argument once already, so when someone raises the point I am not meeting it for the first time in public with an audience watching.

Calibration comes out of the same exercise. Left alone I write "this shows" when the honest phrase is "this is consistent with". I write "because" when I have established correlation and a plausible story. The gap between what I want to say and what I can defend is where most bad analysis lives, including mine, and it turns out to be easy to find that gap when something is actively looking for it.

The style I keep deleting

I do not much like the way AI writes. My particular aversion is signposting: the sentence that turns up to announce that the next sentence matters. This fact is really important. And here is the load-bearing point. Now let us turn to the implications. If a point is load bearing, write it so that it lands and trust the reader to notice. The announcement is what a writer reaches for when they doubt the line can carry itself.

The other habit is flourish. The default register is more decorated than mine. A rhythm where a full stop would do. A third example where two had already made the case. An adjective sent in to do the work of an argument. Left alone it produces prose that sounds like it is being read aloud at a conference.

So a real share of my time goes on deletion. I strip the signposts, flatten the register, cut the third example. It takes longer than people assume and it is not cosmetic, because removing a flourish usually settles whether there was anything underneath it.

Where this leaves me

The model has no stake in whether I am right. I do.

It will not be the one answering when a reader says the chart is misleading, or that the causal story does not hold, or that I have quietly assumed the conclusion. I will be. So I use the tool to find my mistakes before you find them, and when the tool fails, which it does, the mistakes are still mine.

I use AI. The words are mine.