Your Defaults Are Showing

Sometime last year, the em dash became a confession. I used to love a good em dash. But now, I avoid them all, because a reader who spots one stops reading and starts skimming. Machine-made words feel less valuable on instinct, and the writing pays for it before it's really been read. But the em dash is just the beginning. The deeper tells live in the voice.

The writing isn't exactly bad. It's just generic. We were trained to worry about the obvious failures, the ones where AI makes things up or gets a fact wrong. And it does do all that. The harder problem is that it writes well, in a voice that belongs to no one, and once you can hear it you catch it everywhere.

There's a reason it sounds that way. These tools learned to write from us, then got graded on the results. People scored millions of answers, and the ones that scored best were the ones that looked the part: balanced, thorough, wrapped up with a bow. So that's what your AI reaches for now, on reflex, whether the writing needs it or not. The default is generic.

Here are the three to look for:

The first is false enthusiasm. The AI thinks everything is a breakthrough. Every tool is powerful, every idea is exciting, every result is a game-changer. But people aren't actually that impressed that often. When the energy of the writing runs hotter than reality, that's the machine talking. Match the volume to the facts and most of the hype disappears.

The second is the word "thing." Watch how often AI reaches for it: "the thing that matters," "a few things to consider," "here's the thing." The model is playing the odds, and "thing" is the safe bet. It fits everywhere, so it's never quite wrong and never quite anything. The fix is almost mechanical. Every time you see "thing," ask what the real noun is. The decision you keep dodging. The invoice that's three weeks late. If you can't name it, the sentence isn't finished.

The third is the one I almost published without noticing. Call it the over-explanation. The AI gives you a good, concrete example, and then it can't resist adding one more sentence that tells you what the example meant. I recently wrote an issue about using AI as a sounding board, and the initial draft ended a story with this line: "A good board rarely answers the question you walk in with. It finds the sharper one underneath." Sounds like wisdom. But it was also dead weight, because the story right before it had already shown exactly that. I cut it. The example was already doing the work, while the AI just wanted to narrate the moral on top.

You might assume none of this matters, that readers don't notice. They do, and they're getting better at it fast. It turns out the people best at spotting AI writing are the people who use it most. The better you get with AI, the more you become one of them: you start hearing the generic voice everywhere, including in your own drafts.

The internet's answer to all this is to "humanize" your text so it slips past an AI writing detector. The detectors are unreliable, and they flag plenty of real human writing as fake. And that's baked in: the AI learned from good human writers, so the patterns a detector flags as "AI" are often the ones strong writers were already using. There's no clean line to draw.

But more to the point, beating a detector is the wrong goal. The aim shouldn’t be to hide that you used AI; it should be to keep the writing sounding like you.

Issue 5 made a cousin of this point: AI doesn't fail loudly, it fails fluently. It gets things wrong in calm, confident, well-formatted prose. Its voice does the same. It defaults fluently. The fix is the same too. You have to read the output like an editor rather than accept it like a customer.

My process is simple. I keep a checklist of AI tells, and I run my AI against it at least twice for everything I write. First, when I'm generating the initial draft, and then again at the end to check against it. Every time I find a new AI tendency or other writing guideline I’d like to follow, I add it to the list.

I pulled out the ones that work for anyone, cleaned them up, and made them yours: the AI writing tells checklist (it’s free). There are two versions inside. One is for you to read. The other is a file you drop straight into your AI project folder, and it learns your preferences as it goes.

One warning: don't overdo it. The goal isn't zero em dashes and a banned-word list a mile long. I still use an em dash when it earns its place. A tell used once on purpose is fine, and not every sentence is worth the effort. A quick logistics email can stay generic; the attention belongs in the writing that's meant to sound like you. You're just making sure a person is still in there.

Generic is free and everywhere now, which makes it cheap. The part of the writing that sounds like one specific person, paying attention, is the part no machine can hand you.