AI writes a lot of my posts — Editorial Intelligence

Not the ideas. Not the experiences behind them. Not the decision about what I actually believe. But a lot of the words, yes.

AI writes a lot of my posts.

That’s probably the simplest way of describing it.

Not the ideas. Not the experiences behind them. Not the decision about what I actually believe.

But a lot of the words? Yes.

I’ve been thinking about this because there’s an increasingly strange conversation around AI-assisted writing where we’re trying to work out how much AI is too much.

My workflow would probably fail some people’s test.

I have a full-time job. I have a family. I don’t particularly want to spend every evening turning every idea I have into 800 polished words.

So I built a different workflow.

Managing editor, not writer

I think of myself increasingly as the managing editor.

I have the observation.

I decide whether there’s something in it.

I brief the idea into Editorial Intelligence, a system I’ve been building that contains my previous research, arguments, frameworks, published work and half-developed thinking.

AI can then work across that context.

It might find a connection to something I wrote three weeks ago. Challenge the premise. Develop the argument. Produce a first draft. Restructure it. Cut it down for LinkedIn.

In that relationship, AI is doing something surprisingly close to what a writer would do for an editor.

I’m comfortable saying that, because pretending otherwise makes the whole discussion less useful.

I don’t need to have typed every sentence for the underlying thinking to be mine.

The line I’m actually watching

But there is a line, and it isn’t the one people usually argue about.

If AI gives me an idea and I publish it because it sounds clever, I’ve stopped being a managing editor and become an approver.

That’s a different job.

A managing editor briefs, interrogates, rejects, redirects and takes responsibility for what runs. An approver reads something plausible and says yes.

From the outside those two produce exactly the same thing: a published piece with a name on it. Which is why the difference has to be checked from the inside.

So I’ve written the checks into the system rather than leaving them as good intentions:

Where did the idea originate? Name the moment. If the honest answer is that the model suggested it, that isn’t disqualifying — but it changes what the piece is.

Could I defend this argument without AI sitting beside me? In a meeting. To a sceptical colleague. Under follow-up questions.

Is this connected to something I actually observed, experienced, researched or previously thought?

Did AI develop my argument, or quietly give me one?

That last one is the hardest to answer honestly, because a good model will hand you something that feels like a conclusion you reached yourself.

Those questions matter to me considerably more than whether somebody spots an em dash and decides a machine wrote it.

Why detection is the wrong conversation

There are four different things collapsed into the word “wrote”, and they come apart under any pressure at all.

Where the idea came from. What the evidence is and who gathered it. Who decided this was worth saying, in this form, now. And who produced the sentences.

Only the last one is what people are usually testing for.

It’s also the least consequential of the four.

A reader can be entirely correct that AI was involved in a piece of writing and still have established nothing about whether the thinking behind it is any good, original, or the author’s. Detection answers a question about prose. It doesn’t answer the question about authorship that people actually care about.

The unglamorous reason for all of this

Time.

Before AI, the constraint wasn’t ideas. It was the production cost of doing anything with them.

Research it. Find the old notes. Develop the argument. Draft it. Leave it overnight. Rewrite it. Cut it. Format it for somewhere else.

Do that properly and an observation made over breakfast can consume an evening.

Now I can spend more of the limited time I have on the parts where I think I’m actually useful: noticing things, making connections, deciding what’s interesting, challenging what comes back, and deciding what deserves to be published at all.

AI absorbs a lot of the production in between.

What I’d want to be judged on

I should be honest that this arrangement suits me, which is a reason to be suspicious of my own defence of it.

The checks above are mine. Nobody audits them. The whole thing rests on my willingness to answer four uncomfortable questions truthfully when nobody would know if I didn’t, at the end of a long day, about a piece I’d quite like to publish.

That’s a weak enforcement mechanism and I’d rather say so than pretend the system solves it.

But it’s the same weak mechanism that has always governed editorial standards, and it’s roughly what a byline has always meant: not that somebody typed every word, but that somebody is answerable for the result.

That’s probably the most accurate description of what Editorial Intelligence has become for me.

I’m not trying to automate having ideas.

I’m trying to make sure having a life doesn’t prevent me doing something with them.