“Was this AI?”
Three words. No hello, nothing — just that, in my inbox one Tuesday morning last spring, attached to a blog post I’d spent two evenings writing for a client. Two evenings. The first draft was bad — I’d typed half of it standing at the kitchen counter waiting for the kettle, and it read like it. The second evening I argued with one paragraph for twenty minutes, then cut it completely.
So no. It wasn’t AI. It was me, twice, with tea going cold both times.
I told him so, politely. Then I sat looking at my sent email thinking: fine — how would I actually prove that? He’d run the piece through a famous detector; it handed him a percentage. I had feelings and a memory. What settled it was my drafts folder: notes, an outline, the humiliating first version, all timestamped. About a minute. But neither of us had been able to check the thing we actually cared about. He had a number. I had a folder. The truth was in the folder, and the folder was the last place either of us thought to look.
I get versions of that message constantly now. And — awkward part — I send them too, in the other direction: I edit other people’s drafts for a living, and some arrive smelling unmistakably of machine.
What it smells like from the editor’s chair
No single sign convicts anyone — I’ve been the accused, I should know. But when they pile up, four or five of them holding hands, I know what I’m looking at.
The first is emptiness wearing good clothes. Every sentence behaves and there’s no price in it, no date, no name — nothing that could only have come from one real Tuesday in one real place. I edited a piece about invoicing software once that ran eleven hundred words without mentioning a single amount of money. In the margin I wrote: “Did anything ever cost anything in this article?” Never heard back. Swap the keywords and it fits any topic on earth, because it was never about anything. It was about the shape of an article.
Then the rhythm. Or the absence of it. Human paragraphs lurch — a long one while you’re working something out, then two short ones when you land. Generated drafts cruise: topic sentence, three tidy points, little bow on top, next paragraph, same size. I once caught myself squinting at a draft like you check pictures on a wall for level. Everything level. Nobody real writes level.
Repetition is the sneakiest, because it feels substantial while you’re reading. The piece circles its opening claim in fresh sentences, going wider instead of deeper. I summarised one mid-edit and got a single paragraph back. One paragraph, five coats.
And the facts, where I’d put money. A feature the product doesn’t have. A statistic from a study that evaporates when you go looking. I once spent an hour — actual hour, coffee, eleven tabs — chasing a number presented with the serene confidence of a weather forecast. It did not exist. Plausible is the model’s whole job. True is a different job, and nobody assigned it.
Underneath all of it, the deepest sign is simpler: nobody is in the piece. No “I tried this and it broke.” No disagreement with the crowd. No strange small detail that’s only there because it happened. AI has read about everything and experienced nothing, and the absence gets loud. Like a house where nobody cooks. Everything clean. Nothing smells of dinner.
The awkward bit: the detectors are worse than I was
People assume I trust the detectors, given the job. I did — until I ran one of my own blog posts through a big one. A post from 2021, written before any of this was public, about a project I still have the paperwork for. It came back mostly AI. Supposedly.
I sat with that for a while.
My friend writes English as her second language and gets flagged constantly, on work she’s plainly sweated over. Clear, correct, careful English turns out to be exactly what these tools are trained to suspect, which means the more precisely you write in a language that isn’t your first, the more the machine thinks you’re a machine. There’s something genuinely rotten in that.
Meanwhile genuinely generated text sails through whenever a human has edited it with any care. Same article, two detectors, one afternoon: 90% in one, 10% in the other. If two thermometers disagreed like that you’d throw both out and feel the forehead.
So what is a score? A reason to look more closely. That’s the whole honest use of it. Treating it as a verdict has cost students grades and freelancers clients over numbers nobody can defend in a straight line. Half the honest writers I know use AI for an outline or a rough pass, the way you’d use a spellchecker without confessing. The real question was never “was a machine involved?” It’s: is it accurate, is it specific, does it help whoever’s reading it? Bad writing is bad writing, whatever keyboard it came from.
Writing so the question becomes pointless
You don’t out-write a machine by sprinkling typos like seasoning — readers smell that too, faster. You do it by saying what a model structurally cannot say, because it wasn’t there.
Start with the only-you details. The price you actually paid. The date it broke. The customer who complained, loudly, in the group chat, with screenshots. The week-two failure. One honest failure teaches more than three paragraphs of advice, because a failure has an address and advice doesn’t.
I keep two sentences pinned above my desk for this. The generated kind, arriving by the truckload: “Choosing the right project management tool is essential for modern teams looking to streamline workflows and boost productivity in today’s fast-paced environment.” Nothing. Costs nothing, worth about that. And the human kind: “We lasted eleven days on the free plan before the argument started — not about the tool, about whose job it was to update it. Nobody’s job. That’s the review, really: the software was fine, and we were the bug.” Eleven days. An argument. A culprit. None of that grows out of general knowledge, because none of it is general.
Pick a side, too. The models hedge like they’re paid by the hedge — five options, all excellent, all “depending on your needs.” Readers don’t need a catalogue; they need a call. This one, for a beginner, on this budget, because of this — and the famous one is overrated, and here’s why. Reasonable disagreement is close to a fingerprint. Models are trained to avoid it. Humans leave it everywhere.
Bring evidence nobody else has. Interview the one customer. Run the small test and publish your numbers, including the embarrassing one — especially the embarrassing one. Take the photo yourself instead of renting stock pictures.
And edit like your reader has a bus to catch, because they do. Read the draft out loud — in the room, not in your head. Where you run out of breath, the sentence is too long. Where a claim has no example under it, find the example or delete the claim. Where the same point turns up twice, cut the weaker twin and don’t hold a funeral. Say what you tested and what you couldn’t verify, in plain sight. Readers trust that faster than polish. So, these days, does Google.
The desk list, short version: three details only someone who was there would know. Every price and date checked against something real. One recommendation you’d defend in an argument. One piece of evidence that’s actually yours. Hit most of that and no percentage matters. Including mine, that Tuesday.
Questions people actually ask me
What’s the single most reliable sign of AI writing?
There isn’t one.
Can I trust a detector score at all?
As a nudge to read more carefully, yes. As a verdict, no — and I say that as someone nearly convicted by one over a post written at a kitchen counter. They flag the innocent, spare the guilty, and disagree with each other about the same paragraph on the same day. If a score is the only evidence against a piece of writing, there isn’t any evidence; there’s a number, and a person who’d rather trust the number than do the reading. I keep my drafts now — notes, outlines, the awful first versions — and anyone who writes for money should, because a process unfolding messily across two evenings outweighs any percentage ever quoted at me. The folder settled it. Keep the folder.
Will AI-assisted content rank on Google?
Yes — when it’s accurate, specific, and genuinely useful to whoever lands on it. Google’s fight is with mass-produced filler built to game rankings, not with whichever drafting tool you used. Checked facts and real editing were the difference long before any of this existed.
Should I just avoid AI completely, then?
No. Research, structure, a rough first pass — normal tools, increasingly unavoidable, fine. The line is publishing raw output nobody checked. Everything before that line is a tool. Everything after it is a shortcut, and shortcuts are exactly what readers have learned to smell.
Last thing
You can often sense AI writing. You can rarely prove it. Be suspicious of anyone — person, software, percentage — who claims they can, to a decimal point.
So aim at the better target. Write with prices and dates. Admit the failure. Recommend one thing and put your name behind it. Readers aren’t short of content anymore; they’re drowning in it. What they’re short of is writing by somebody who was actually there.