Is my AI showing?
I let AI help me think out loud, then didn't like how it sounded
Confession time, though it's not much of one, because I don't think anyone reading regularly was fooled: the last few posts I've written about AI — the team-shape one, the judgement one, the whole "Whose Future Is It Anyway?" run — were drafted with heavy AI assistance1.
Not "AI wrote them and I hit publish." More that I was trying to build a proper thesis in real time, out loud, across several posts, and the fastest way I found to drag the shape of an idea out of my own head was to argue it out with ChatGPT first and edit the result after.
It worked, in the sense that the thesis is genuinely mine and I stand behind it. It also produced some of the most oddly flat writing I've put on this blog in a very long time of doing this2, and I couldn't immediately tell you why, because I was too close to it.
So — because apparently my solution to "did AI change how I sound" was to ask another AI — I had Claude read every post on this blog, from 2009 through to now, and rate how much each one sounded like AI wrote it, then explain its reasoning.
I fully expected to be told that em-dashes, bullet points, and "it's not X, it's Y" were the smoking gun. Everyone's told that. It's the entire "did a bot write this LinkedIn post" genre at this point.
Instead I got told, essentially: no, calm down, you've been doing that since before ChatGPT existed.
The stuff that isn't actually a tell, for me
It turns out that my first ever post on this blog, from 2009, builds its argument with a numbered list. I used the "it's not X, it's Y" reframe in a piece about voice assistants in 2016, and again in some agency blog copy in 2018, both comfortably pre-dating any large language model anyone had heard of. I've been using bold pull-quotes as a kind of visual full-stop since at least 2015. And the em-dash thing — I'd already written a whole post back in June 20253 pointing out that people (like me) who've used different typographic dashes correctly for twenty years are now getting side-eyed for it. It's even a caveat on the "About me" page, for crying out loud!
None of that is AI creeping into my writing. It's just my writing. Lists, repetition-for-effect, a dash instead of a comma when I want the reader to pause slightly longer than a comma allows — that's not a tell, that's me doing words much gooder than what other people does them, innit.
Which meant the actual answer to "what changed" was somewhere else, and it turned out to be more interesting than "he used too many em-dashes."
The stuff that actually is a tell
Density, not presence
I've always reached for the odd rhetorical trick, but sparingly — once or twice a post, at a moment that's earned it. In the AI-heavy run, the same device shows up five or six times in a single piece, because that's roughly the rate at which a model reaches for it when nothing's stopping it.
Specificity, or the lack of it
A post I write unassisted has a name in it, or a number, or a memory that's embarrassingly precise4. The AI-scaffolded posts trade almost entirely in abstractions — "capability", "agency", "the system"5 — because a model has no memories of its own to reach for, and I hadn't yet done the work of putting mine back in.
Being opinionated, unashamedly
I think it's also the absence of anything that could get me in trouble.
No swearing. No opinion stated in a way that risks someone disagreeing with me personally rather than academically. No typo left in because I noticed it and thought the joke was funnier than the correction.
My unassisted writing has always had a bit of character and a sense of having skin in the game, because I am proud of having strong opinions (albeit loosely held). The scaffolded stuff, by contrast, reads like it was drafted by someone extremely reasonable, talented, knowledgeable, and experienced who has never lost an argument at a dinner party, which is not a description anyone who actually knows me would apply.6
None of which makes those posts wrong — the thesis inside them holds up, and I still believe every word of the argument. It just means the words weren't spoken authentically in my voice.
So I'm speaking up again
Once I could see what the actual gap was, the fix was obvious, if not exactly quick: pull the run of related posts back together into one properly built essay — same thesis, same four-forces framework I'd been circling around post-by-post, but written the way I actually write, with the specifics and the opinions and the odd bit of nonsense put back in.
It's now up in the thinking section of the site as its own thing, separate from the blog, because it's a slightly different kind of object: "The Real Effect of AI: From My Desk to Everyone Else's Future". The same argument you'll have seen me building here in pieces over the last few weeks, but it sounds much more like me.
If you've read this far, you've probably clocked a few of the tricks I've just described happening in this very post, deployed on purpose, at the density I'd normally use them. That's not an accident. Consider it the control group.